Intelligent overflights for the safety of vulnerable users

When conducting road safety analyses, it is crucial to have a detailed understanding of the movements of the entire network under analysis. Certain parameters are more likely to reveal potential safety issues, and are almost systematically analyzed in safety studies. These key elements include, for example, the origins and destinations of trips, the routes traveled, the speed used, the quantification of flows, or the slots available.

In the past, conducting surveys to obtain these different parameters was quite expensive. Indeed, such surveys required simultaneous observations at several points of the network by different people, and numerous data processing afterwards.

The FNX-INNOV team decided to put these technologies to the test in the context of a safety study for the sector of the Polyvalente des Monts, a study conducted with the City of Sainte-Agathe-des-Monts and the Centre de services scolaires des Laurentides. The first step of the study was to establish a diagnosis of the current situation with regard to safety, both on the school site and in the adjacent sector.

The following paragraphs summarize the project team’s steps to evaluate the drone survey technology, the methodology considered and the conclusions of these processes.

Methodology

The proposed methodology for the technology evaluation was established by the project team after conducting several small-scale tests.

The identification of the analysis area must include significant origins and destinations for the trips under study. The school site was completed with adjacent streets to include areas used for drop-off or parking (by parents or employees), pedestrian crossings near the site, and the school bus drop-off. Thus defined, the area selected for the analysis of the Polyvalente is a rectangle of approximately 300m by 180m.

Given the limitations of flight time for a light drone (20-30 min), it is necessary to target the analysis periods. Generally, the most significant periods are directly related to traffic flows (large flows or highly conflicting flows). A traffic count can help to quickly identify these periods. For the school project, these periods were easily identified, as they are the intervals corresponding to the beginning and end of classes, i.e. between 8:40 and 9:00 am and between 4:40 and 5:00 pm.

For the video recording, the project team used a Mavic 2 Pro drone, hovering at an altitude of 100 m, with the camera oriented at a 35 degree angle. The recording was done with 4K quality and a frame rate of 30 frames/s.

It should be noted that the use of drones is strictly regulated in Canada. In order to fly a drone, it is required to have a pilot’s license, to reserve the airspace, to ask for an authorization for the use of the land for takeoff and landing, to respect specific safety measures, etc.

The preliminary processing of the videos was done automatically with the TrafficSurvey application from the company DataFromSky to extract the raw data and prepare the analysis. To identify and classify moving objects (vehicles, pedestrians, etc.), each object is isolated by comparing successive images in the video. Once the objects are identified, the software assigns them a position for each frame. The automatic analysis of the successive positions is then translated into data on direction, speed and acceleration. The size of the objects, combined with the dynamic movement characteristics, allows a reliable classification by cars, pedestrians, bikes, motorcycles, trucks, buses, etc.

Semi-automatic processing is also possible in the case of videos that do not have a good resolution. For the presented project, this additional processing was necessary to ensure 100% detection for some schoolchildren that were only partially detected by the initial processing, i.e. schoolchildren at more than 200 m.

At the end of this step, the visual representation of the data already provides very useful information for the safety diagnosis (see figure 2, in red the pedestrian paths, in green the cars and in purple the buses). It is thus easy to identify areas of vehicle-pedestrian conflict, cases of non-compliance with traffic regulations or dangerous maneuvers, for example.

• Automatic processing of raw data to obtain intersection flows and paths taken between intersections – By graphically configuring virtual “gates” on the reference image, the software can quickly provide all information about the objects that passed through each pair of two “gates.” This analysis module not only allows for quick counts at intersections, but also provides the origin-destination matrices in the analysis area. In addition, other information is easily obtained through this processing, such as travel times between two “gates” or stopping times.
• Automatic processing of raw data to identify possible near misses, i.e. objects whose trajectories, speeds and/or accelerations could have led to a collision.
• Automatic processing to calculate available slots in traffic flows (not evaluated in this project).

Information was extracted to supplement the count data, to identify car maneuvers and stops before the end of classes, and to qualify the relative importance of student trips and bus trips, among other things.

It should be noted that the interpretation of the data is not necessarily different from a project without the benefit of drone video analysis. The principles governing safety studies remain the same, however, the use of artificial intelligence and partial automation of data processing offers significantly more relevant information to professionals in this field.

In the present project, several analyses have been significantly simplified by the use of aerial video data:

• The identification of significant areas of car-pedestrian conflict and the subsequent treatment of these areas.
• The identification of significant areas of car-pedestrian conflict and the subsequent treatment of these areas.
• The identification of significant areas of car-pedestrian conflict and the subsequent treatment of these areas.

Conclusions and perspectives

The use of aerial videos in conjunction with AI software opens up access to detailed information on mobility dynamics, information that is otherwise much more difficult to collect by conventional methods. It should be mentioned, however, that the use of drones has limitations for the time being, particularly with respect to continuous flight time (maximum 40 minutes for the best models), very strict legal oversight, and to some extent the dependence of the survey on acceptable weather conditions. However, the flight time can be extended, for some types of UAVs, by using a tethering system (powering the UAV by cable).

Regarding the AI software evaluated, the limitations are mainly related to the quality of the video images. Successive tests have allowed us to recommend recording at 4K, at a height of 60 m at an angle of maximum 30 degrees from the vertical. These conditions allow the software to detect and classify movements well, without having to resort to additional corrections.

Finally, the visuals resulting from the automatic data analysis with the AI software also become powerful supports for the presentation of the results of the studies, simplifying the understanding of the identified problems.

Following the evaluation on several projects during the year, FNX-INNOV’s Transport and Smart Mobility department has decided to introduce the practice of using aerial videos and image processing by AI software on all projects that can benefit from it and to optimize the processing methodology for each type of project.