In March 2022, Fondazione Innovazione Urbana and the Municipality of Bologna built a new public space for children in Via Procaccini (Bologna) near a middle school (see Figure 1), using the approach of tactical urban planning and participatory design. The area chosen to build the new square was the subject of a mobility study conducted by Systematica and Transform Transport. The objective of the study was to monitor pedestrian and vehicular flows, with the aim of detailing specific patterns of space use. The monitoring was carried out through observations supported by a camera and video analytics techniques. The research produced a series of analyses related to observed flows in the area during the pre/ post-intervention phases. This was meant to support the iterative design process based on the tactical urbanism approach.

Introduction

Planning infrastructures and services for sustainable urban mobility is one of the main challenges for European cities (Buhrmann, S., Wefering, F., Rupprecht, S., 2019), which are increasingly facing problems of traffic congestion, road safety, energy dependence, and air pollution (United Nations, 2016). In this context, planning efforts are focusing on walkability (Abley & Hill, 2005; Speck, 2013), namely how friendly the urban environment is for walking or spending time in public spaces. Improving the level of walkability of an urban area implies the presence of a comfortable, safe, and barrier-free pedestrian infrastructure, as well as people-friendly environments that allow opportunities for social interaction (Gehl, 2013).

The focus on walkability, which began with the principles highlighted in the European Charter of Pedestrian Rights issued by the European Parliament in 1981, has become even more evident given the unprecedented effects of the Covid-19 pandemic on urban mobility (Transform Transport, 2022). In July 2020, the European Commission (European Platform on Sustainable Urban Mobility Plans, 2020) provided guidelines for the implementation of pedestrian mobility planning interventions (e.g., extension of sidewalks, queue management in transportation infrastructure, creation of temporary public spaces, etc.) aimed at ensuring that citizens can access basic services on foot (i.e., 15 Minute Cities) (Moreno et al., 2021).

Although traditional approaches to the study of pedestrian mobility tend to focus on the spatial dimension (Annunziata & Garau, 2020; Steinfeld, 2011), the individual characteristics of pedestrians have a significant impact on the perceived level of walkability. As highlighted by the 2030 Agenda for Sustainable Development adopted by the United Nations Member States (United Nations, 2016) (i.e., SDG 11.2-Sustainable Transport for All), urban mobility should be designed to be more inclusive with respect to the needs of those in vulnerable situations, such as the elderly, children, and people with disabilities.

In this context, the research focuses on evaluating walkability for children aged 5-13 years, mainly middle school students. The study is part of recent efforts in designing safe, comfortable, and livable streets and public spaces for children (Peyton, 2019; Aerts, 2018; Danenberg, Doumpa & Karssenberg, 2018). Indeed, walkability for children includes the opportunity to play freely in the open air, walk independently and safely, and develop a sense of belonging to their neighborhood (Krysiak, 2020).

The study was conducted by Systematica and Transform Transport, in collaboration with the Fondazione Innovazione Urbana and the Department of Urban Planning, Housing and Environment of the Municipality of Bologna, with the aim of supporting the planning of sustainable and inclusive mobility strategies in the city of Bologna (see Figure 2). The first phase of the research (i.e., macroscopic scale) began in December 2020 with the objective of assessing the level of pedestrian accessibility of public services dedicated to children’s needs through GIS analysis of georeferenced data made available by the Geoportal of the city of Bologna. The results of this phase of the research were then presented on May 13, 2021, at the 6^th Biennale dello Spazio Pubblico (Abdelfattah et al., 2021).

The second phase of the research (May to September 2021) focused on assessing the level of walkability for children in the city of Bologna, according to the following criteria: (i) accessibility of services; (ii) comfort of pedestrian infrastructure; (iii) road safety; and (iv) attractiveness of public spaces. The results of this GIS analysis (Gorrini et al., 2022) focused on the Navile neighborhood (i.e., mesoscale), selected by the Municipality of Bologna for the design of a new public space for children.

The third phase of the study (February to April 2022) involved Fondazione Innovazione Urbana and the Municipality of Bologna for the design of a new square (i.e., microscale), which was implemented in March 2022 through the approach of tactical urbanism and participatory design as part of the EN-UAC project “EX-TRA – EXperimenting with city streets to TRAnsform urban mobility” (No. 99950032). In particular, the new square was built in an area used for unregulated parking located in Via Procaccini (Navile), nearby a middle school.

The area was the subject of the study conducted by Systematica and Transform Transport, aimed at monitoring pedestrian and vehicular flows during the pre and post-intervention phases. The monitoring was carried out through camera-supported observations and automated image analysis techniques (Ibrahim, Haworth and Cheng, 2020; Barthélemy et al., 2019; Zhao et al., 2019). The study produced a series of temporal and spatial analyses related to the observed flows in the area (e.g., density maps, space use, average speeds, road crossing behaviors, etc.), with the aim of supporting the iterative design process based on the tactical urbanism approach.

Enabling Data and Methodology

Data Collection Planning

As part of the redevelopment of the area in Via Procaccini, a video data collection campaign was carried out with the aim of conducting a quantitative analysis of the use of the space during the pre- and post-intervention phases. Specifically, the data collection campaign involved monitoring the area for about two months, as follows:

• Pre-intervention: from 07/02/2022 to 04/03/2022;
• Post-intervention: from 02/04/2022 to 31/04/2022.

Video data collection was focused on three daily time slots, with the aim of analyzing space use and activation characteristics in relation to:

• 7 am – 9 am: school entrance;
• 1 pm – 3 pm: school exit;
• 5 pm – 7 pm: spontaneous use of the area.

Data Collection Tools and Placement

Video data were collected using an EXIR camera with a wide-angle lens, an image quality equal to 1280×720 pixels, and a frame rate equal to 15 FPS (Frames Per Second).

The placement of video data collection equipment was guided by principles of flexibility and reproducibility of the study related to the possibility of relocating the equipment to new areas of intervention. In addition, it was determined by the need to obtain an overall shot of the space to determine its uses (see Figure 3).

Following these principles, the camera was installed on a pole at the corner of Via Procaccini and Via di Vincenzo (north side), at a height of about 3m and with a mean axis of framing parallel to Via Procaccini facing the area of intervention.

Data Analysis Process

The data collected by the camera were processed in four stages:

1. Recognize and track objects in the videos using deep learning algorithms;
2. Transpose the recognized objects from perspective view to 2D view through a process of geo-referencing the frames;
3. Process the geo-referenced data and develop summary metrics for characterizing and quantifying the recognized object events in the area;
4. Visualize and map the developed metrics using GIS.

Object Recognition and Tracking

The first phase of the analysis was carried out using deep learning algorithms for object recognition and tracking to quantify pedestrians and vehicles in the area. In particular, two open source algorithms were used: YOLO v5 and Deep Sort.

YOLO-You Only Look Once (Jocher et al., 2021; Gutta, 2021) is an object recognition algorithm based on convolutional neural networks (i.e., CNN). In the context of this study, an open source model of Yolo v5 was used, trained on images taken from CCTV cameras in the city of Montreal.

Deep Sort is an object tracking method based on a recursive application of the Kalman filter, which allows a unique ID to be assigned to the same object recognized in multiple images. More specifically, the Kalman filter uses estimated information about an object’s position and velocity to make predictions about its future location.

The result of the object recognition and tracking process is the creation of a text document in which information about the objects is made explicit in an anonymous form. Each line of the document describes the video frame, object ID (expressed in numeric characters), object class (i.e., pedestrian, vehicle), pixel X, and pixel Y.

Geo-referencing

The second phase of the analysis focused on translating the image coordinates related to the objects identified in the videos (i.e., pixel X, pixel Y) into geographic coordinates (i.e., latitude, longitude). The objective of this phase is to eliminate perspective distortion of the images by locating pedestrians in the target area.

The methodology uses the QGIS implementation of the Thin Plate Spline algorithm as a transformation technique, with Ground Control Points selected to anchor points in the perspective image. The procedure is repeated through an automatic matching algorithm in each frame of the video.

Data Preprocessing and Metrics Development

The third phase of analysis included data validation/ preprocessing and the development of summary metrics designed to quantify and characterize the observed area. In this phase, data from individual videos were merged and spatially discretized on a 1m x 1m grid.

The merged data were then filtered, with the aim of mitigating the influence of factors that systematically prevent object recognition in certain areas of space (i.e., under-sampling) or that promote repeated object recognition (i.e., over-sampling). These may be related to geometric framing configurations or weather and lighting conditions.

To this end, cells that had a cumulative value of recognized objects below the first percentile or above the ninety-ninth percentile of the data distribution were filtered out of the analysis at each time level.

In addition, for the analyses on average pedestrian speeds observed in the area, an additional data validation process was carried out by filtering out objects with a speed greater than the ninety-eighth percentile from the distribution of values. 

GIS Analysis and Mapping

The fourth phase of the analysis focused on the production of maps and graphics obtained through the open-source software QGIS v3.26. Maps describing the average use of the area observed in the phase before and after the tactical urbanism intervention were elaborated (see Figure 4).

Data Analysis Metrics

Data metrics were then implemented, describing the use of the new public space in Via Procaccini. These are discretized spatially and temporally as follows:

• Spatial discretization aims at aggregating pedestrian patterns in the study area on a 1m x 1m grid with the following classification of cells:
  • Public Space:
    • Games;
    • Seating;
    • Benches.
• Temporal discretization aims at identifying cumulative and granular use patterns with a logic of progressive disaggregation such as:
  • Average pre-/ post-intervention days;
  • Weekdays/holidays;
  • Time slots (7:00-9:00 am, 1:00-3:00 pm, 5:00-7:00 pm);
  • 15-minute intervals.

Metrics

Cumulative Permanence Time

Cumulative permanence time, expressed in equivalent minutes, quantifies the total number of minutes spent in the study area, discretized on a 1m x 1m grid. This metric made it possible to identify the most heavily trampled areas, hierarchizing the intervention area in relation to its uses. Cumulative dwell time is calculated as the sum of detected pedestrians in a cell, cumulated at different temporal disaggregation.

Average Speed

Analyses of average speeds have made it possible to describe how pedestrians and vehicles use urban space, providing spatial and behavioral classification into:

• Places of rest and play;
• Crosswalk areas;
• Areas of vehicular circulation.

Average speed is a cell-based metric, calculated as the mean velocity in a cell at different temporal disaggregation. Velocities were computed through the Python library MovingPandas which enabled the extraction of trajectories’ attributes for the detected pedestrians.

Results

Times of Activation

The first phase of analysis focused on the study of activation times of Via Procaccini in the phases before and after the tactical urbanism intervention. This analysis focused on weekdays and holidays, divided into three time slots identified as moments characterized by entry to school (7:00 am – 9:00 am), exit from school (1:00 pm – 3:00 pm), and spontaneous use of the area (5:00 pm – 7:00 pm). The three time slots were then divided into 15-minute intervals with the aim of characterizing at a granular level the patterns of use and frequentation of the area.

The level of use was quantified as cumulative dwell time, expressed in equivalent minutes  with the aim of estimating the number of minutes spent in the study area. More specifically, equivalent minutes correspond to the sum of minutes spent on each cell of the 1m x 1m grid onto which the area is discretized at each time interval.

Charts 1 and 2 summarize the results of the analysis for weekdays and holidays, depicting the cumulative occupancy time in the period before and after the redevelopment intervention, and the percentage difference in the time intervals. The cumulative occupancy time increases in most of the time intervals following the intervention, confirming the transition from a crossing area to a rest and play area.

Chart 1 shows the trends for weekdays. Significant percentage use level differences are shown between 7:30 am – 7:45 am (+87%) and between 2:15 pm – 2:30 pm (+88%). These results denote a tendency to stop more in the area during the times before and after school hours. In addition, there is a strong percentage increase in the 5:00 pm – 7:00 pm time slot (+77%), in which the area, following the redevelopment, is activated in relation to dynamics no longer related to school functions.

Chart 2 summarizes the results of the analysis for holidays, highlighting patterns of use not related to school functions. Growth peaks in cumulative dwell time emerge, reaching +270% in the 1 pm – 3 pm range. There is also an increase in dwell times in the afternoon slot, with an average increase of +33%. In contrast, the percentage differences related to the morning describe an increase in use related to minimal dwell times and may be related to the seasonality of data collection.

Space Utilization

The space utilization of the new public space in Via Procaccini was quantified through metrics describing cumulative dwell time and average speeds. The cumulative dwell time, expressed in equivalent minutes, made it possible to quantify the number of minutes spent in the study area discretized on a 1m x 1m grid and identify the most trampled areas, thus hierarchizing the intervention area in relation to its uses and describing the overall change in space use by its users.

In particular, the analysis of pedestrian flows was aimed at quantifying the presence of pedestrians and the new spatial configurations defined because of the redevelopment. The results show 43% growth in the cumulative dwell time recorded in the area following the redevelopment intervention. Strong changes in the spatial distribution of pedestrians in the study area are also highlighted.

Figure 5A shows the pedestrian use patterns of Via Procaccini in the phase before the redevelopment of the area. Two main passage corridors are highlighted at the two sidewalks on Via Procaccini and a parking area on the sidewalk between Via di Vincenzo and Via Procaccini. Three crossing corridors complementary to the crosswalks between Via Procaccini and Via di Vincenzo are also noted, corresponding to the extension of the sidewalks (1, 2) and the diagonal visual axis of the area (3).

Figure 5B represents the new pedestrian dynamics that originated because of the redevelopment. There is a general increase in dwell time (+43%) and a shift in the preferred place for stopping and playing, which is now identified in the new public space.

Natural crossing corridors (1, 2 in Figure 5A), integrated into the redevelopment design, are activated again; new ones also emerge (1, 2, 3 in Figure 5B). 1, 2 are related to the placement of benches in public space, which blocks the diagonal axis of the area, lengthening the diagonal crossing (1) and promoting passing phenomena adjacent to the square on the side of the roadway (2). Finally, the crossing area at the crosswalk between Via Procaccini and Via di Vincenzo (3) widens, including diagonal flows connecting the square and the school.

In addition, the analysis of average speeds aims to characterize the new modes of use of public space. Figure 6 (A and B) presents the average pedestrian speeds recorded in the study area during the pre- and post-intervention phases. In general, there is a decrease in speeds throughout the area, with a sharp decrease in the new public space. There are also higher speeds in the areas corresponding to the roadway in both phases of the study.

Finally, an increase in average speed at the crosswalk between Via Procaccini and Via da Faenza is shown in Figure 6B, followed by a decrease at the sidewalk on Via Procaccini.

Public Space

Figure 7 shows the spatial configurations of the new public space originated from the tactical urbanism intervention. The cumulative dwell time of the area, transformed from a crossing area to a rest and play area, increases by 216%. It is also noticeable (see Figure 7B) that the crossing corridors in the area, which were used before the intervention, remain active. However, new use characteristics emerge related to the present furniture (1, 2 games, 3, 4 seats, 5, 6 benches), which have the most significant cumulative dwell times in the area.

More specifically, Table 1 lists summary metrics describing the use of the three types of furniture in relation to the overall use of the new public space during the post-intervention phase. Seating has the largest increase in average permanence time compared to the new public space (+100%), followed by benches (+82%) and games (+79%).

Chart 3 shows the delta of use of the furniture with respect to the overall space of the new public space in the three time slots of analysis. Strong percentage growth in the use of the three objects can be seen on weekdays, where benches and seating show the largest changes, particularly in the afternoon (benches +203%, seating +195%, games +163%). On holidays, on the other hand, an inverse trend emerges in the morning, where the use of the area related to crossing needs prevails. In the daytime and afternoon, on the other hand, strong positive deltas emerge for benches (13-15 +176%, 17-19 +64%) and games (13-15 +116%, 17-19 +37%).

Figure 8 presents the spatial characteristics related to average speeds in the new public space. This presents a major lowering of average speeds (-42%), related to the new play and parking functions within it. Four crossing corridors in the area, complementary to the play areas identified in the furniture analysis, are highlighted, where the average speeds present values above 0.82 m/s.

Final Remarks and Future Work

Results made it possible to quantify through a critical reading the effectiveness of the urban regeneration intervention implemented in Via Procaccini. The proposed data-driven design approach is presented in this case through the monitoring of the pre/ post-intervention phases.

As shown in Figure 9, space use is in direct relation to the form and functions present. Previous analyses, in fact, have highlighted the delta between pre/ post-intervention phases and the role that furniture plays in influencing dwell time. By analyzing the maps with a critical view, it is possible to identify the successful and critical elements of this intervention and use these observations to inform subsequent design phases.

In the case of cumulative dwell time, which measures the use of the different areas of the new public space, the flows and infrastructure do not always coincide. The crosswalks across Via Procaccini are not perceived by the flows as the only possible place to cross the street. In particular, the crosswalks at the height of Via di Vincenzo are only partially used by pedestrians crossing the street; most pedestrians use the space in direct connection with the centerline of the street, cutting diagonally across the roadway. In parallel, the crosswalks at the height of Via da Faenza are not particularly used. In contrast, the crossings parallel to Via Procaccini are used by most users, as is the sidewalk opposite the public space, which manages to effectively channel flows.

In addition to the crossings outside the crosswalks, it is also possible to identify a flow parallel to the new public space, outside the planters, which follows the prevailing direction of travel in the widening. This alternative route occurs within the roadway and often in the opposite direction of travel, remaining outside the impassable furnishings of the pedestrianized space.

The elements of success and criticality highlighted in Figure 9 contain information that is extremely useful to a possible subsequent design phase. The shift from a quantitative approach, based on the analysis of collected data, and a synthetic/ design approach is the basis of data-driven design: the goal is to maximize the effect of positive interventions and resolve the critical issues found. The ideal approach to this type of design is the construction of an iterative process, which quantifies each intervention separately and whose design part is repeated after each monitoring phase, creating a continuous circle of hypotheses and solutions.

---

The results of this research have been presented at the 22^nd International Walk21 Conference on Walking and Liveable Communities, 19-23 September 2022, Dublin (Ireland).

Acknowledgment

We thank Fondazione Innovazione Urbana and the Municipality of Bologna for their fruitful collaboration. An extended Italian version of the study report is available at this link. The analyzed data were treated according to the GDPR-General Data Protection Regulation (EU, 2016/679). This research didn’t receive any specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

References

Abdelfattah, L., Boni, G., Carnevalini, G., Choubassi, R., Gorrini, A., Messa, F., Presicce, D. (2021). A user-centric approach to the 15-minute city. Examining children’s walkability in Bologna. In: Proceedings of the 57th ISOCARP World Planning Congress 2021, 8-11 November 2021, Doha (Qatar), pp. 582–592. Available at: https://doha2021.dryfta.com/programme/discussions/sponsor-lounge/proceedings

Abley, S., Hill, E. (2005). Designing living streets-a guide to creating lively, walkable neighbourhoods. Living Streets.

Aerts, J. (2018). Shaping urbanization for children: A handbook on child responsive urban planning. UNICEF. Available at: https://www.unicef.org/media/47616/file/UNICEF_Shaping_urbanization_for_children_handbook_2018.pdf

Annunziata, A., Garau, C. (2020). A literature review on walkability and its theoretical framework. emerging perspectives for research developments. In: International Conference on Computational Science and Its Applications, Springer. pp. 422–437. https://doi.org/10.1007/978-3-030-58820-5_32

Barthélemy, J., Verstaevel, N., Forehead, H., & Perez, P. (2019). Edge-Computing Video Analytics for Real-Time Traffic Monitoring in a Smart City. Sensors, 19(9), 2048. https://doi.org/10.3390/s1

Buhrmann, S., Wefering, F., Rupprecht, S. (2019). Guidelines for Developing and implementing a sustainable urban mobility plan – 2nd edition. Rupprecht Consult-Forschung und Beratung GmbH. Available at: https://www.eltis.org/mobility-plans/sump-guidelines

City of Montreal (2021). Annotated images taken from the video stream of traffic cameras. Information Technology Department. Available at: https://donnees.montreal.ca/ville-de-montreal/images-annotees-cameras-circulation

Danenberg, R., Doumpa, V., Karssenberg, H. (2018). The City at Eye Level. STIPO Publishing. Available at: https://thecityateyelevel.com/app/uploads/2019/06/eBook_CAEL_Kids_Book_Design_Kidsgecomprimeerd.pdf

European Platform on Sustainable Urban Mobility Plans (2020). COVID-19 SUMP Practitioner Briefing. CIVITAS SATELLITE CSA. Available at: https://www.eltis.org/sites/default/files/covid-19_sumppractitionersbriefing_final.pdf

Gehl, J. (2013). Cities for people. Island press.

Gorrini, A., Presicce, D., Messa, F., Choubassi, R. (2022, submitted). Walkability for Children in Bologna: Beyond the 15-minute City Framework. Journal of Urban Mobility.

Gutta, S. (2021). Object Detection Algorithm — YOLO v5 Architecture. Medium. Available at: https://medium.com/analytics-vidhya/object-detection-algorithm-yolo-v5-architecture-89e0a35472ef                                                                                                                            

Ibrahim, M. R., Haworth, J., & Cheng, T. (2020). Understanding cities with machine eyes: A review of deep computer vision in urban analytics. Cities, 96, 102481. https://doi:10.1016/j.cities.2019.102481

Jocher, G., Chaurasia, A., Stoken, A., et al. (2021). ultralytics/yolov5: v5.0 – YOLOv5-P6 1280 models, AWS, Supervise.ly and YouTube integrations (v5.0). Zenodo. https://doi.org/10.5281/zenodo.4679653

Krysiak, N. (2020). Designing Child-Friendly High Density Neighbourhoods. Transforming our cities for the health, wellbeing and happiness of children. Cities For Play. Available at: https://www.citiesforplay.com/_files/ugd/534edb_5e8553bb853d40228da3083a0ed1eede.pdf

Moreno, C., Allam, Z., Chabaud, D., Gall, C., Pratlong, F. (2021). Introducing the “15-Minute City”: Sustainability, Resilience and Place Identity in Future Post-Pandemic Cities. Smart Cities, 4, 93–111. https://doi.org/10.3390/smartcities40100066

Peyton, A. (2019). Designing Streets for Kids. NACTO-Global Designing Cities Initiative. Available at: https://globaldesigningcities.org/publication/designing-streets-for-kids/

Speck, J. (2013). Walkable city: How downtown can save America, one step at a time. Macmillan.

Steinfeld, E. (2011). Universal Design Handbook – Second Edition. McGraw-Hill Publishers.

Transform Transport (2022). Shifting Paradigm: the impact of Covid-19 on transport planning (second edition). Fondazione Transform Transport ETS. Available at: https://issuu.com/systematica/docs/shifting_paradigm_2_transform_transport

United Nations (2016). Transforming Our World: The 2030 Agenda for Sustainable Development. United Nations. Available at: https://sustainabledevelopment.un.org/post2015/transformingourworld/publication

Zhao, Z., Zheng, P., Xu, S., & Wu, X. (2019). Object Detection with Deep Learning: A Review. IEEE Transactions on Neural Networks and Learning Systems, 30, 3212-3232. https://doi.org/10.48550/arXiv.1807.05511

Since the introduction of the Social Logic of Space (Hillier and Hanson, 1984) and the further developments of Space Syntax theories during the following decades (Hillier et al., 1996; Hillier, 2007), the proposed methodologies have been proven effective to analyze the space in his physical configuration. In this regard, the study proposes the application of Visibility Graph Analysis (VGA) measures to Piazza Duomo, one of the busiest public squares in Milan, Italy. These metrics were tested in relation to the pedestrian movements measured through remote sensing technologies. Deep-learning video-analytics methods were used to detect the pedestrian occupancy in five different moments of the day (see Figure 1). Detections were then reprojected on the plane to obtain the spatial utilization, discretized on the grid used to calculate the VGA metrics. Then, a series of correlation analyses tested the grid-based analyses against the cumulated footfall and a movement metric. The results show how the correlation between the VGA measures and the pedestrian movements varies greatly at different radii, considered through the variation of the restricted visibility parameter, ranging from negative weak correlations to positive moderate correlations for the cumulative density, and from positive moderate correlations to weak correlations for the turnover metric. The results, in line with other recent research works, shows how pure VGA measures can only partially describe an open public space, posing an interesting step in the direction of future works, aimed at the development of a more complex model based on the Space Syntax theories.

Literature Review

Space Syntax theories have been tested since their introduction in the effort to correlate the syntactic measures with empirical evidence, such as the recorded flows of pedestrians or vehicles. In the case of the Grid-based analysis (i.e., VGA), the metrics were tested in various settings. In one of the most notable example, an urban environment case study is represented by the experiments made on the Barnsbury area in London (Turner, 2003), based on 116 gate counted pedestrian movements. More recently, a building-based case study is used as setting for an occupancy analysis in an educational building (Tomé et al., 2015). In this case, video analytics is used to identify moving and standing individuals, to analyze the utilization patterns of the space.

Empirical pedestrian gathering techniques can be divided in two main categories: automatic and manual systems (Conroy, 2001). Both can rely on sensing technologies to quantify space utilization, as outlined in Table 1 (van Nes et al., 2021). Among them, videos represent the most versatile mean of gathering, as they can be mounted at a fixed location (i.e., snapshots, general movement traces, gate counts, mean occupancy), or mobile (i.e., pedestrian following).

Table 2 classifies the technologies presented in Table 1 based on automatic data counting methodologies, purposes and limitations and Space Syntax analyses typologies. Data counting methodologies can be divided in two categories:

• Algorithm-based counting methods, namely those that rely on additional analyses to retrieve data (i.e., images, videos);
• Sensor-based counting methods, which depend on the technology embedded in the sensor to retrieve data (i.e., Wi-Fi, Bluetooth, GPS/ LBS technologies).

The analysis purposes categorization, see Table 2, is structured in three groups, enabling the study of:

• Spatial patterns, pedestrian behaviors, and movements in an area (e.g., grouping, sitting, etc.);
• Densities and the quantification of the footfall around hotspots;
• Routes and the analysis of routing preferences (e.g., most walked paths, etc.).

Limitations of sensors can be classified as:

• The analysis is restricted to a limited area, as for images and videos (i.e. extension);
• The analysis is restricted to a limited number of people, as for Wi-Fi, Bluetooth, GPS technologies (i.e., users).

Lastly, the sensors are classified based on which Space Syntax analyses can be linked to them:

• Line-based analyses relate to route preferences and traces (e.g., videos, GPS/LBS);
• Grid-based analyses mainly relate to densities and patterns (e.g., images, videos, Wi-Fi, Bluetooth).

Methodology

Case study definition and VGA metrics selection

Piazza Duomo (Milan, Italy), is chosen as a case study as it is structured as a pedestrian space, interrupted by a limited number of elements, as the subway entrances. In the analyses, the square is studied together with the area surrounding it, with a buffer of 250 meters from the center of the square, to create an uninterrupted linear system. The drawing used in Depthmap is obtained from a collection of open geodata, which have been enriched to include pedestrian spaces not visible from satellite. Figure 2 shows the extent of the pedestrian are considered for the analysis.

The analyzed space is configured as an open-geometry area, that is expected to be used in diverse manners, as follows: (i) users crossing the central area to get to the other side of the square; (ii) users walking slowly and stopping to appreciate the architecture of the square; and (iii) users walking in relation to urban functions (e.g., subway accesses, shopping venues, etc.). VGA metrics chosen to describe the area are listed in Table 3, these are calculated with diverse restricted visibility distances (i.e., NR – Not Restricted, 150m, 100m and 50m).

Video description and analytics techniques

Five videos were analyzed to obtain pedestrian footfall data. The footage was recorded on July 15, 2021, in five different moments of the day, specifically at 08:00, 11:00, 12:45, 15:00 and 18:00, each of them depicted a 30-minute time interval, with a size of 1920×1080 pixels and a frame rate of 15 frame per second (FPS).

Yolov5 (Jocher et al., 2021) with DeepSORT integration was used to detect pedestrians in the square, as visible in Figure 3. An open-source model was used, which was trained on webcam images in Montreal, Canada (City of Montreal, 2021), obtaining a mean Average Precision (mAP) equal to 0.809 in pedestrians’ recognition on the original training set.

Video georeferencing

A georeferencing technique was implemented, to estimate the geographical position of the distribution of pedestrians in the square. The methodology uses the QGIS implementation of the Thin Plate Spline algorithm as a transformation technique, with 240 Ground Control Points selected to anchor points in the perspective image. The procedure is repeated through an automatic matching algorithm in each frame of the videos.

Detections’ measurements

The georeferenced pedestrian coordinates (x, y) in each frame were then associated to the corresponding cell of the VGA grid, in Figure 2. Then, a mean occupancy value, expressed as people/sec, was computed, averaging the number of detections that fell into the same grid cell within fifteen frames. Finally, two metrics were computed:

• Cumulated footfall measure, defined as the cumulative sum of the mean occupancy value in thirty seconds bins;
• Turnover variation measure, computed as the sum of the difference in occupancy in a cell every five seconds and the activation measure (i.e., measuring the frequency of activation of the cell), normalized on the mean occupancy of the cell.

Results

Measured pedestrian footfall

The video analytics process allowed for the calculation of the occupancy of the space on the 2x2m grid. Since the footage was taken at different hours of the day, the footfall measured in the videos differs in magnitude and patterns, reflecting the varying nature of the public space. Figure 1 shows the detection collected during the 08:00 video, and Table 4 shows the results of the video analytics, including descriptive statistics to highlight the diversity of the results.

Then, outliers were removed from the distributions, considering the 1st and the 99th percentile of each video sequence and two metrics were defined:

• Footfall Data (FD) is defined as the sum of the cumulated mean occupancy for the five videos. It represents the utilization pattern of the square, describing an average situation of different moments of the day (see Figure 4);
• Movement Data (MD) is defined as the mean value of the measure of variation (Var) for the five videos. It represents the cells with a high turnover, namely the areas where the pedestrians don’t stop but generally keep moving (see Figure 5).

VGA metrics

The VGA analyses were calculated for the area shown in Figure 2, for several maximum visibility distances (i.e., NR – Not Restricted, 150m, 100m and 50m), to compare the results and understand which distance can describe effectively an open public space. In Figure 6, it is possible to see how out of the ten selected metrics, three of them (i.e., I_Cm – Isovist Compactness, I_DM – Isovist Drift Magnitude, and I_O – Isovist Occlusivity) are not related to the maximum visibility distance parameters, since those are based on the un-restricted isovist’s shape. The other metrics show a consistent difference among the four distances, namely the transition from the large-scale topology to the open public space topology.

Correlation analyses

In this phase, the VGA metrics, with the variation of the restricted visibility distance parameter, and the Footfall Data (FD) were normalized, while the Movement Data (MD) was linearized. Then, using Pearson’s Correlation Coefficient, the datasets were analyzed with a series of correlations analyses. The results for the Footfall Data (FD) and the Movement Data (MD) are shown, respectively in Figure 7 and Figure 8.

Discussion

The correlation results for the Footfall Data (FD), visible in Figure 7, do not show strong correlations between the analyzed data. The values range from negative to positive correlations, but the absolute values are never higher than .376, in the case of the Isovist Drift Magnitude. The restricted visibility parameter influences results for (i) Connectivity, (ii) Through Vision, (iii) Visual Control, and (iv) Visual Integration (HH).

The correlation analysis results for the Movement Data (MD), shown in Figure 8 express different outcomes. The correlation coefficient is still moderate at best, such as the cases of Connectivity (.373 – NR and .411 – 150m), Through Vision (.373 – NR), Visual Clustering Coefficient (-.326 – NR and -.391 – 150m), Visual Control (.402 – NR and .468 – 150m), and Visual Integration (HH) (.396 – NR and .303 – 150m)

These results can be related to a case-specific utilization of the space for the Duomo case study. The touristic nature of the square and the presence of the subway accesses lead to a higher cumulative density in the middle part of the space, but, at the same time, the area utilized to move around draw some “desire lines”, especially near the gallery, that are also well represented in the VGA analyses. The Movement Data (MD) metric allow to clean the recorded movement data from the noise of people standing still in the space, narrowing the analysis to the actual flows.

In support of this statement, it is possible to see how the short-radius Visual Integration (50m), which shows a positive correlation of .326 with the Footfall Data (FD), is not linked to the topological shape of the larger environment, but it is very similar to results obtainable in a perfect square shape: higher values in the middle, with lower values toward the edges, also the Connectivity follow a very similar pattern. Meanwhile, the Movement Data (MD) is better correlated with the large-radius and, thus, showing how the environment shapes the presence of pedestrian flows. Finally, in the case of Through Vision, which can be regarded as an alternative to EVA agents for 360° field of view, the comparison between the Footfall Data (FD) and Movement Data (MD) shows much more consistent results in the latter case.

The outcomes of these analyses are in line with other recent research studies (Ericson et al., 2020; Koutsolampros et al., 2019), and show limitations in the application of VGA metrics in relation to different typologies of spaces. However, the proposed case study is a particularly complex environment chosen to stress the reliability of the VGA metrics in a space characterized by an inherent asymmetry of functions and the presence of strong attractors and generators, namely the subway accesses. These factors weight heavily in the pedestrian movements, greatly influencing the cumulated density as expressed in the FD dataset. On the other hand, the utilization of a “movement-based” metric, expressed in the MD dataset, allow to partially re-focus the analysis on the movements.

Conclusions and Future Work

The proposed case study is utilized to test the VGA metrics values, obtained by varying the restrict visibility parameter, in the central part of Piazza Duomo (Milan, Italy). The central portion of the public space was recorded through a webcam, and the footage was studied through video analytics techniques, measuring the pedestrian activity of the space discretized in a 2x2m grid.

The metrics and the pedestrian movements were tested in correlation analyses, showing weak to moderate correlation values. In this framework, the proposed Movement Data (MD) shows positive moderate correlation with VGA metrics usually associated to movements, namely Connectivity, Through Vision and Visual Integration (HH). As future improvement, this dataset can be refined by introducing tracking technologies in the video analytics techniques, associating a unique ID to each pedestrian, and recording its location and movement speed.

Future works include the application of the Agent Analysis (Turner, 2003), based on the VGA metrics, considering a series of access gates, acting as starting points of space exploration and the implementation of anchors-based metrics, specifically a metric value representing the distance from the main anchors. The weighting of the VGA metric, especially the Visual Integration (HH) with the distance from selected points could lead to a more reliable representation of the public space, in the effort to build a calibrated VGA model.

Lastly, future work will be focused on the possibility to further characterized pedestrian profiles by using proxemics, speeds and trajectories data collected through tracking video analytics techniques, in line with the conceptualization of different walking behaviors of single pedestrians and groups in (i) time driven pedestrians, (ii) space driven pedestrians and (iii) social driven pedestrians.

---

The results of this research work have been presented at the 13th International Space Syntax Symposium: Messa, F., Ceccarelli, G., Gorrini, A., Presicce, D., Choubassi, R. (2022). Deep Learning Video Analytics to Assess VGA Measures and Proxemic Behaviour in Public Spaces. In: Proceedings of the 13th International Space Syntax Symposium (13SSS), 22-24 June 2022, Bergen (Norway). Available at: https://www.hvl.no/globalassets/hvl-internett/arrangement/2022/13sss/479messa.pdf

Acknowledgments

The analyzed data were treated according to the GDPR-General Data Protection Regulation (EU, 2016/679). This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

References

City of Montreal (2021). Annotated images taken from the video stream of traffic cameras. Information Technology Department. Available at: https://donnees.montreal.ca/ville-de-montreal/images-annotees-cameras-circulation

Conroy, R. (2001). Spatial navigation in immersive virtual environments. University of London, University College London (United Kingdom).

Ericson, J. D., Chrastil, E. R., & Warren, W. H. (2020). Space syntax visibility graph analysis is not robust to changes in spatial and temporal resolution. Environment and Planning B: Urban Analytics and City Science, 48(6), 1478-1494. https://doi.org/10.1177/2399808319897624

Hillier, B. (2007). Space is the machine: a configurational theory of architecture. Space Syntax.

Hillier, B., & Hanson, J. (1984). The Social Logic of Space. Cambridge, New York: Cambridge University Press.

Hillier, B., Major, M.D., Desyllas, J. Karimi, K. Campos, B., Stonor, T. (1996). Tate Gallery, Millbank: a study of the existing layout and new masterplan proposal. UCL (University College London): London, UK. Available at: https://discovery.ucl.ac.uk/id/eprint/932/1/hillier-etal-1996a_TateGallery.pdf

Jocher G., Stoken A., Chaurasia A., Borovec J., NanoCode012, TaoXie, Kwon Y., Michael K., Changyu L., Fang J., Abhiram V., Laughin, tkianai, xyNONG, Skalski P., Hogan A., Nadar J., imyhxy, Mammana L., … & wanghaoyang0106 (2021). ultralytics/yolov5: v6.0 – YOLOv5n ‘Nano’ models, Roboflow integration, TensorFlow export, OpenCV DNN support (v6.0). Zenodo. https://doi.org/10.5281/zenodo.5563715

Koutsolampros, P., Sailer, K., Varoudis, T., & Haslem, R. (2019). Dissecting Visibility Graph Analysis: The metrics and their role in understanding workplace human behaviour. In: Proceedings of the 12th International Space Syntax Symposium (Vol. 12). Available at: https://discovery.ucl.ac.uk/id/eprint/10073528/

Tomé, A., Kuipers, M., Pinheiro, T., Nunes, M., & Heitor, T. (2015). Space–use analysis through computer vision. Automation in Construction, 57, 80-97. https://doi.org/10.1016/j.autcon.2015.04.013

Turner, A. (2003). ‘Analysing the visual dynamics of spatial morphology.’ Environment and Planning B: Planning and Design, 30(5), 657-676. https://doi.org/10.1068/b12962

van Nes, A., & Yamu, C. (2021). Empirical Data Collection and Analysis, and Connecting Data with Space Syntax. In A. van Nes, & C. Yamu (Eds.), Introduction to Space Syntax in Urban Studies (pp. 133-170). Springer. https://doi.org/10.1007/978-3-030-59140-3_5

Covid-19 pandemic has deeply affected urban mobility: the social-distancing strategies adopted to cope with the virus transmission pushed the majority of city users into avoiding public transport services in favour of safe and contactless travel options. To investigate this phenomenon, this research proposes an Urban Informatics approach to understand passengers’ opinions and expressed polarity through user-generated social-media data (Twitter) in the Greater London Area. The analysed corpus consists of 27,700 tweets posted between January 13th, 2020 and May 17th, 2021, and geolocated in immediate proximity of a short sample of selected Tube stations. Data is segmented in several phases, based on the restriction measures enforced by the UK National Authorities. Each subset is then analysed through texts structuring and semantic analyses, with LDA topic modelling and Sentiment Analysis. Finally, the outcomes of these processes are interpreted in their chronological succession and compared to demand data and service-disruption data.

User-generated Social-media Data as a Tool for Understanding the City

Data is at the foundation of modern planning practice and among other tools that enable collection of mobility patterns and flows. In particular, user-generated social-media data has the potential to represent both national-scale (Jurdak et al., 2015) and urban-scale mobility patterns (Plunz et al., 2019).

This opportunity has become even more crucial considering the need to investigate the unprecedented effects of disruption of the Covid-19 pandemic on urban mobility. In particular, the nation-wide lockdown and post-lockdown phases have drastically changed citizens behaviours and mobility patterns related to the use of public transport services (Buhrmann et al., 2020), due to the need to avoid crowded transport infrastructures in favour of safe and contactless travel options (e.g., private vehicles, cycling, walking, etc.).

In this context, the article presents an extended analysis of a Twitter-data corpus, collected by geographical location and specific timeframes, and further filtered by removing non-relevant tweets and non-human posting. The acquired corpus consists of 27,700 tweets geolocated in immediate proximity of selected Tube stations of the Greater London Area (UK) and posted between January 13th, 2020 and May 17th, 2021. The social-media data is segmented in 9 phases, shown in Figure 1, to compare the outcomes of the analyses in timeframes that have a mobility-demand significance, but also a social connotation.

The goal of these analyses is to offer a comprehensive view of the relation between public transport and user-generated data. Hence, the process is twofold: firstly, the analysis is focused on understanding the impact of the pandemic on the public transport mobility patterns through daily demand data and daily service-disruption data. Secondly, the social-media data is investigated with two different methodologies based on texts structuring and semantic analyses, specifically: (i) identification of tweets’ subjects through LDA Topic Modelling, and (ii) assessment of the conversation emotional polarity through Sentiment Analysis.

Social Media Data Gathering

In accordance with the research scope, the phasing structure is derived from the evolution of the pandemic; the phases were selected by considering the national and local policies imposing restrictions affecting Greater London. The calendar, comprising of 25 policy milestones, was reduced to 9 phases in order to avoid an excessive segmentation of the outputs, considering only the major self-distancing measure changes. In addition to the pandemic period, a Baseline Phase was added for the purpose of providing a reference scenario for the assessed indicators (see Figure 2). This practice is common in a before-after analysis and the timeframe chosen was in line with the Coronavirus publications produced by Transport for London (TfL).

Similarly, a selection of Tube stations (see Figure 3) was performed in order to identify the locations to look for geo-located tweets. This process was achieved by using demand data from 2019 published by TfL, which represents the daily entry/exit for the typical weekday of a pre-pandemic year. The stations were selected by considering the 85th percentile of the demand, resulting in 40 stations located in the Greater London Area.

The selected stations’ locations are used as centroids to calculate buffers of 400 meters of radius. The measure of the proposed catchment area around stations was chosen since it can be considered as a 5-minute walking distance in the analysis setting. Moreover, the small catchment area represents a location bias able to represent the content posted in proximity of public transport infrastructure.

The raw outcome of this process is a dataset comprising of 95,463 tweets, which was then post-processed in several steps in order to harmonize tweets’ text to be input in further semantic analyses, removing noise and superfluous information, by:

• Detecting repetitive and non-human posts in the overall corpus, the presence of which could lead to biased and inaccurate topic modelling;
• Text content pre-processing with characters-based simplifications removing punctuation signs, emojis, URLs and converting all the text to lowercase format;
• Natural Language Processing (NLP) methodologies, such as stop words removal and lemmatizations (Bird et al., 2009).

All these processes contribute to the harmonization of text and the removal of words with no semantic meaning, enhancing the performance of the topic modelling, specifically: (i) stop words removal was based on the NLTK English stop-words list (e.g. a, an, the, etc.) (Bird et al., 2009), which was extended with custom words (e.g., London, etc.) that appeared ubiquitously in the corpus; (ii) lemmatization consists of a reduction of each word to its root and allows a further selection of allowed tokens, in this case leading to the selection of nouns, verbs, adjectives and adverbs in each tweet. The data cleaning process was differentiated for topic modelling and sentiment analysis and it’s shown in Figure 4.

Text Semantic Analysis

The two analyses performed to comprehend the content of the corpora were (see Figure 5):

• Topic modelling: to gain knowledge on the main tweets’ subjects
• Sentiment analysis: to investigate the expressed message polarity.

Topic Models are statistical language models used to understand the theme structure of a document. In particular, the used model was Latent Dirichlet Allocation (LDA), which is a probabilistic generative model that takes as input a collection of lemmatized texts and outputs topics’ probability distribution over each document (Kapadia, 2019). For each phase of the analysis timeframe, topic modelling was used to identify 10 topics per period.

The model parameters were hypertuned to obtain best fitting topics, looping over LDA’s main parameters. Subsequently, the significance of the results was quantified using the topics’ coherence metric, which measures the degree of semantic similarity between high-scoring words in the topic, helping to distinguish artifacts from semantically valid outcomes. In order to measure the influence of pandemic-related posts in the analysed corpora, a reference control set built around Covid-19 related words was defined and compared against the previously obtained topics.

By using the reference set as an unseen document to query the models, the output was the probabilistic distribution of the queried control words among the phases’ topics set. Furthermore, Sentiment Analysis was performed using Vader (Valence Aware Dictionary and sEntiment Reasoner) (Hutto et al., 2014), a lexicon and rule-based analysis tool developed to estimate the polarity of social media texts. The tool outputs three values divided into positive, neutral and negative scorings and a compound value to summarize the sentiment’s results.

Topic Modelling Analysis

Since the topic modelling analysis was performed on diverse corpora and the LDA model is generative, the reference was devised to infer the relation between Covid-19 related terms during the phases.

The set consists of 17 terms that also include variation of the same lemma, divided thematically into five categories:

• Pandemic: covid, corona, coronavirus, virus, pandemic;
• Lockdown: lockdown, quarantine, staysafe, stayhome;
• Prevention: distance, distancing, mask, facemask;
• Vaccine: vaccine, covidvaccine, vaccination;
• Other: nhs.

This set was built considering the overall corpus content by manually evaluating a random subset of tweets, Figure 1 shows the distribution of these words.

The relation between the reference set and the phases’ corpora is expressed by measuring how many terms of the set fall into the dominant topic (i.e., the topic with highest semantic similarity with the queried text) and by the overall topic weight (i.e., the probability distribution of the topics with respect to the queried text).

Demand Data and Tweets

In addition to the demand used to select the stations, which was referred to the pre-pandemic year (2019), the daily demand of the selected timeframe was analysed in order to understand the actual impact of Covid-19 on the Tube’s mobility. The variation between the phases is substantial to the extent that during the last phase (P8) the average daily demand is -67% compared to the Baseline Phase (see Figure 6).

Service Disruption Analysis

In order to structure a comprehensive view of the public transport usage during the pandemic evolution, the service disruption was taken into consideration. In this regard, the disruptions caused by diverse sources were categorized and aggregated per phase. The service interruptions and delays were considered for all the Tube lines since the selected stations are distributed among all the subway lines and a service quality index was calculated in term of minutes of delay per day, as shown in Figure 7.

Sentiment Anylysis

While the Topic Modelling Analysis is performed separately on each corpus, the Sentiment Analysis is used to interpret each single tweet, then the results are aggregated and interpreted per phase. VADER tool expresses the polarity of each item by three values, defined as positive, neutral and negative and a normalized weighted Compound Score to summarize the overall results (Hutto et al., 2014). For the scope of this research the value considered is the Compound Score, which expresses a single comparable score among the entries.

Findings

While the demand and service data describe a currently well-known phenomenon, the text semantic analysis allows the gathering of more insights on the perception of the pandemic. The distribution of covid-related terms in the different phases allow to understand the evolution of the conversation: during the Lockdown 1 phase, the conversation was more focused on the actual situation, with the largest part of the conversation focused on the “Pandemic” theme. This was followed by an interest more related to the “Lockdown” theme and the restriction measures, during Lockdown 2 phase. Finally, the discussion converged towards the “Vaccine” theme during the last analysed phases.

By comparing these considerations to the topic modelling outputs, it is possible to confirm some of these insights. During the Lockdown 1 phase (P1), the dominant topic relates to “Pandemic” and “Lockdown” themes and it shifts toward “Lockdown” in P4 and to “Vaccine” during the last phases. However, the Dominant topic weight metric shows how relevant the topic is in relation to the generated model. A more or less relevant topic can be influenced by several factors, such as an effective topic modelling process or the semantic significance of the analysed corpus. However, for the scope of this research, it can be considered as a reference-set independent weight metric describing the relevance of the outputs.

Finally, the sentiment analysis can be interpreted using the threshold proposed by the tool creators (Hutto et al., 2014): typical compound limits are positive (score ≥ 0.05), neutral ( – 0.05 > score > 0.05) and negative (score ≥ – 0.05). Considering these values, all the phases are characterized by a positive average compound score. However, social-media data is usually influenced by a positivity bias (Waterloo et al., 2018) because negative emotions are often considered as not appropriate to publicly share on social media. For this reason, the trend metrics are proposed, and it is possible to appreciate how the general sentiment per phase is always lower than the Baseline Phase and especially how the variation closely follows the demand data, hence the restrictions implementation calendar.

Even if in a qualitative fashion, these sets of data manage to add information that is otherwise difficult to obtain without the utilization of surveys or time-consuming and cost-impactive methods. For this reason, the analysis of social media data is steadily increasing its applications, ranging from sociological and medical studies to mobility-related analyses. Crowd-sourced data could be one of the potential data feeds needed to develop a comprehensive model of an urbanized area (digital twin) able to describe and simulate phenomena happening in the urban environment, because in addition to location-based information, they allow to identify social and cultural patterns.

---

The results of this research work has been presented at the 49th European Transport Conference 2021 (ETC 2021, 13-15 September 2021 – Online), organized by AET, and published on ZENODO: Messa, F., Ceccarelli, G., Gorrini, A., Presicce, D., Deponte, D. (2021). Analysing ‘Twitter Conversation’ of London Tube Stations: The Case of the COVID19 Pandemic. In: Proceedings of the 49th European Transport Conference 2021 (ETC 2021), 13-15 September 2021 – Online. https://doi.org/10.5281/zenodo.6493631

References

Bird, S., Klein, E., & Loper, E. (2009). Natural language processing with Python: analyzing text with the natural language toolkit. O’Reilly Media, Inc.

Buhrmann, S., Wefering, F., Rupprecht, S. (2019). Guidelines for Developing and implementing a sustainable urban mobility plan – 2nd edition. Rupprecht Consult-Forschung und Beratung GmbH. Available at: https://www.eltis.org/mobility-plans/sump-guidelines

Hutto, C., & Gilbert, E. (2014). Vader: A parsimonious rule-based model for sentiment analysis of social media text. In: Proceedings of the International AAAI Conference on Web and Social Media 8 (1). Available at: https://ojs.aaai.org/index.php/ICWSM/article/view/14550

Jurdak, R., Zhao, K., Liu, J., AbouJaoude, M., Cameron, M., & Newth, D. (2015). Understanding human mobility from Twitter. PloS one, 10 (7), e0131469. https://doi.org/10.1371/journal.pone.0131469

Kapadia, S. (2019). Topic Modeling in Python: Latent Dirichlet Allocation (LDA). Available at: https://towardsdatascience.com/end-to-end-topic-modeling-in-pythonlatent-dirichlet-al location-lda-35ce4ed6b3e0.

Plunz, R. A., Zhou, Y., Vintimilla, M. I. C., Mckeown, K., Yu, T., Uguccioni, L., & Sutto, M. P. (2019). Twitter sentiment in New York City parks as measure of well-being. Landscape and urban planning, 189, 235-246. https://doi.org/10.1016/j.landurbplan.2019.04.024

Waterloo, S. F., Baumgartner, S. E., Peter, J., & Valkenburg, P. M. (2018). Norms of online expressions of emotion: Comparing Facebook, Twitter, Instagram, and WhatsApp. new media & society, 20 (5), 1813-1831. https://doi.org/10.1177/1461444817707349

The development of new technologies is shaping the growth of cities in many ways. Among these, the Internet of Things (IoT), Artificial Intelligence (AI), the high-resolution global positioning system (GPS), big data and new building materials and techniques are expected to transform cities’ core functioning elements, affecting all aspects of our lives (Joint Research Centre, 2019; Geospatial Commission, 2020). When technology is combined with the growing availability of open data, new possibilities emerge to develop a novel understanding of the urban fabric and use.

Deep learning algorithms are a subset of machine learning algorithms that process and combine their input in ever growing abstractions to obtain meaningful outputs. This project focuses on the analysis of pre-trained deep learning models for object detection, image segmentation and crowd counting and on their applications on research topics linked to the understanding of temporal patterns related to the development of Covid-19, climate analysis and thermal comfort, and their integration with transport modelling.

Deep learning algorithms are of particular interest in the field of computer vision, enabling the manipulation of large datasets in an automatic way. In the past years, the rising availability of deep learning techniques led to new frontiers in the automatic understanding of images and estimating the number of objects within an image. Deep learning methods obtained state-of-the-art results for image classification, object detection (Zhao et al., 2019) and instance/semantic segmentation tasks. Furthermore, in the past years, the availability of pre-trained weights for deep learning algorithms has grown.

Dataset and First Steps

The study area centers around Corso Buenos Aires, Milan, one of the longest and most well-known retail streets in Europe (Transform Transport, 2018). The dataset consists of a collection of images (from MilanoCam) depicting the street with a strong perspective view, showing the flow of vehicles and people from Piazza Lima to Porta Venezia. These are taken at one-hour intervals, with a high potential for comparative analyses of changes in the street use behavior at peak times and different seasons.

The current study builds on previous research presented in Shifting Paradigm 1st edition, which was structured in two phases. In the first phase, a set of algorithms with pre-trained weights were selected based on their popularity and tested on a selection of 16 images. The results were then measured against the manual counts performed on each image. The goal was to define the usability of these models out-of-the-box, for the given dataset. Table 1 outlines algorithms tested within Phase 1 of the research and Table 2 lists the mean deviation between automatic counts and manual ones for each algorithm, computed as:

In the second step, the analysis of the images for a year-long period was carried out. The initial focus was on the performance of the algorithms with respect to weather and light conditions as well as business activity of the street.

Finally, the study aimed to identify the patterns in the street use every day at 19:00, the moment when people leave their offices, with particular attention to the Covid-19 lockdown period. Figure 1 outlines the main temporal trends and patterns for each mode.

Data and Tools

In the last phase, an extensive dataset of 14,956 images was constructed, representing 12 daily timeslots (6am – 10pm) between January 2019 and May 2021. These were then analyzed on the fly using Yolov5 (You Only Look Once) (Jocher et al., 2021) as an object detection algorithm. The goal was to recognize pedestrians and locate them in an image.

YOLOv5 represents an evolution of the YOLO architecture, which enables faster and more accurate class predictions. This architecture relies on CSPDarknet53 as the backbone and PAN as an aggregator level. Furthermore, it uses image augmentation methods in training to improve accuracy in results (Gutta, 2021).

An open-source model was used to detect objects in the images, which was trained on CCTV cameras in Montreal. The dataset represents vehicles, pedestrians, constructions, cyclists, and bus instances, obtaining a mean Average Precision (mAP) equal to 0.809 in pedestrians’ recognition on the original training set (City of Montreal, 2021). Figure 2 shows results on a sample image of Corso Buenos Aires.

Analyses

Building on knowledge obtained from previous research, the study was then structured in four steps, which focus on the analysis of pedestrian patterns in Corso Buenos Aires with increasing temporal granularity (i.e., months, day types, hours) and spatial accuracy (i.e., whole image, sidewalks).

First, an exploratory analysis was performed on the dataset (see Figure 3). The goal was to visualize overall trends in registered pedestrian footfall in 2019. Figure 3 shows average monthly profiles for weekdays and weekends. These present expected patterns, with anomalies in June and September, when extraordinary events (i.e., Fashion week, Giro d’Italia, etc.) occurred in the street. Then, the internal validity of the dataset was assessed from two perspectives.

The first one aimed at quantifying the deviations of the model’s performance for different weather and lighting conditions, identifying systematic inaccuracies in the detections. Here, a manual validation methodology was implemented on 10% of all images, comparing the amount of automatically detected pedestrians with those counted manually. Results shown in Figure 4 demonstrate how darkness and rain worsen overall accuracy, leading to a mean absolute difference with manual counts above 10%. Instead, shadow and sun conditions affect results slightly, leading to a mean absolute difference equal to roughly 7.5%.

The second validity check aimed at defining the reliability of the dataset through a sensitivity analysis on data robustness and outliers. Data for the year 2019 was analyzed to ensure consistency in the analysis. First, a mean hourly number of detected pedestrians was computed. This should be the baseline value for comparisons. Then, hourly detections were grouped in random clusters. For each, the mean hourly value was computed to assess its deviation from the mean value. Figure 5 shows how splitting 2019 data into 5 random groups leads to a deviation from the mean lower than 5% for all moments of the day, with an exception for 7 and 8 am. This may be due to differences in footfall patterns during the weekday and weekends. Instead, splitting data into 5 random groups leads to a higher deviation from the mean, above the 5% threshold at multiple moments throughout the day.

The second phase of the study investigated specific patterns in the use of Corso Buenos Aires. The dataset was divided into five subsets, considering the Covid-19 pandemic restrictions. Then, hourly trends for workdays and weekends and for different day types were studied and compared among different restriction phases (see Figure 6). This analysis aimed at extending the one carried out in the first phase, with a focus on post lockdown restrictions. These can be classified as “necessity” restrictions (namely, movements were limited to necessity reasons, as work and grocery shopping), and “soft” restrictions (namely, leisure movements were allowed, such as shopping). The goal was to determine whether differences among the two phases could be recognized by means of pedestrian counting. Figure 6 gives an overview of the restriction phases and of the overall footfall trends.

Figure 7 outlines detected pedestrian volume trends by day type and hour. The top graph shows how mean pedestrian counts never reached pre-covid levels during the study period, with slight differences registered between “necessity” and “soft” restrictions phases. The bottom graph depicts hourly patterns for each restriction phase. Lower volumes were registered at lunch time and in the evening in “necessity” phases, while higher volumes can be seen in the weekends during “soft” phases. Furthermore, the effects of the curfew implementation in the city at 10pm can be seen from “necessity 2” to “necessity 4” phases.

The third step of the study consists in converting instantaneous pedestrian densities captured in images into pedestrian flow and related walking speed. To do so, pedestrian fundamental diagrams, specifically derived from an innovative application of dynamic microscopic model, were applied on the detections. The Social Force Model represents pedestrians as particles subject to forces (i.e., attraction and repulsion) in an analogy with fluid dynamics, quantifying pedestrian flows under business-as-usual conditions. The Social Distancing Model is a modification of the default parameters to simulate social distancing conditions (Espitia et al., 2022).

Figure 8 outlines the preliminary results on the application of these models to determine the overall influence of social distancing in the study area. The graph depicts Social Force Model and Social Distancing Model curves (red and blue curves), which quantify the hourly pedestrian flow in Corso Buenos Aires, given increasing pedestrian densities (x axis). These can be compared with the occurrences of the detected densities in the image dataset (black curve).

As can be seen, in 52% of the images in the dataset, the number of detected pedestrians would lead to a difference between business-as-usual and social distancing hourly flows greater than 8%. While images cannot be used to compute hourly flow, the graph gives insights on possible flows scenarios in different crowding conditions in Corso Buenos Aires, enabling a visual comparison with the occurrences of detected densities in images.

Lastly, a preliminary analysis on pedestrians’ distribution on sideways was carried out. As Corso Buenos Aires has a north-south orientation, its shading conditions change gradually during the day, allowing for a thorough pedestrian comfort study. In this phase, temperature and cloud conditions were researched by using an open-source meteorological dataset. These may influence the sidewalk choice. Figure 9 shows the shading conditions in Corso Buenos Aires on August 6th, 2021, between 12am and 5pm. The left sidewalk (east) is in the shade from 12am to 2pm, while the right one (west) is shaded from 3pm to 6pm.

In this phase, images were masked to detect pedestrians on the sidewalk of interest. Then, the share of pedestrians for each sidewalk was computed to obtain a comparable metric. Figure 10 outlines trends in sidewalk choice in 2019, depicting the average monthly share of pedestrians per sidewalk and the mean detected temperature. While a general preference for the left sidewalk can be observed, curves abruptly change between 2 and 3 pm in summer months with temperatures above 30° Celsius. These patterns are consistent with 2020 ones, proving potential for further analyses (see Figure 11).

Conclusions and Future Research

This research aims at studying different methods to identify and count objects in images. The goal is to understand the potential of each method with available pre-trained weights. Then, the purpose is to analyse the street use of Corso Buenos Aires by means of ongoing studies.

The evolution in results obtained between the first steps and phase three highlight possibilities in development of new applications and studies involving image datasets. In fact, while models that were trained on broad scope datasets, i.e., COCO (Lin et al., 2014) would lead to modest results for small instances (i.e., pedestrians and cyclists), results showed a great improvement using a model specifically trained on the classes of interest (i.e., vehicles, pedestrians, cyclists) and on images with similar orientation. Further improvements could be obtained through data augmentation, limiting the effects of meteorological factors on images quality, and by fine-tuning the open-source models.

In general, the methodology proves to be time and cost-efficient with respect to the level of accuracy and details inferred while guaranteeing the privacy of individuals (Bernas et al., 2018). The use of an extensive dataset led to statistically significant results on aggregated temporal periods, overcoming the uncertainties of a single image. Furthermore, it gives the possibility to detect meaningful moments to be further analyzed by means of video analytics.

Future works include the growing integration of modelling outputs with real data, the replication and forecast of complex pedestrian patterns at both urban district and city levels, integrating multiple datasets and deepening the study on pedestrian comfort in different environmental conditions. An example is represented by the usage of this information and other data from different sources (wi-fi hot spots, mobile and telco data, etc.) to implement innovative predictive models able to replicate and forecast complex pedestrian flow patterns at both urban district and city levels, as implemented in the city of Melbourne (City of Melbourne, 2021) and in the city of Liverpool, Australia (Barthélemy et al., 2019).

---

Acknowledgments

The analyzed data were treated according to the GDPR-General Data Protection Regulation (EU, 2016/679). This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

References and Online Sources

Barthélemy, J., Verstaevel, N., Forehead, H., & Perez, P. (2019). Edge-Computing Video Analytics for Real-Time Traffic Monitoring in a Smart City. Sensors, 19(9), 2048. https://doi.org/10.3390/s19092048

Bernas, M., Płaczek, B., Korski, W., Loska, P., Smyła, J., & Szymała, P. (2018). A Survey and Comparison of Low-Cost Sensing Technologies for Road Traffic Monitoring. Sensors, 18(10), 3243. https://doi.org/10.3390/s18103243

City of Melbourne (2021). Pedestrian Counting System. Available at: https://www.melbourne.vic.gov.au/about-melbourne/research-and-statistics/city-population/Pages/pedestrian-counting-system.aspx

City of Montreal (2021). Annotated images taken from the video stream of traffic cameras. Information Technology Department. Available at: https://donnees.montreal.ca/ville-de-montreal/images-annotees-cameras-circulation

Espitia, E., Gorrini, A., Vacca, A., Deponte, D., Sarvi, M. (2022). How Covid-19 is Affecting Pedestrian Modelling and Simulation: The Case of Venice. Transportation Research Record. https://doi.org/10.1177/03611981221088224

Joint Research Centre (2019). The future of cities: opportunities, challenges and the way forward. European Commission Publications Office. https://data.europa.eu/doi/10.2760/375209

Geospatial Commission (2020). Unlocking the power of location. The UK’s Geospatial Strategy, 2020 to 2025. Cabinet Office. Available at: https://www.gov.uk/government/publications/unlocking-the-power-of-locationthe-uks-geospatial-strategy/unlocking-the-power-of-location-the-uks-geospatial-strategy-2020-to-2025

Gutta, S. (2021). Object Detection Algorithm — YOLO v5 Architecture. Medium. Available at: https://medium.com/analytics-vidhya/object-detection-algorithm-yolo-v5-architecture-89e0a35472ef

Jocher, G., Chaurasia, A., Stoken, A., et al. (2021). ultralytics/yolov5: v5.0 – YOLOv5-P6 1280 models, AWS, Supervise.ly and YouTube integrations (v5.0). Zenodo. https://doi.org/10.5281/zenodo.4679653

Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C.L. (2014). Microsoft COCO: Common Objects in Context. ECCV. https://doi.org/10.48550/arXiv.1405.0312

Transform Transport (2018). Re-imagining Corso Buenos Aires. Systematica Srl. Available at: https://research.systematica.net/research/re-imagining-corso-buenos-aires/

Zhao, Z., Zheng, P., Xu, S., & Wu, X. (2019). Object Detection With Deep Learning: A Review. IEEE Transactions on Neural Networks and Learning Systems, 30, 3212-3232. https://doi.org/10.48550/arXiv.1807.05511