ALGORITHMS FOR MEASURING SAFETY BETWEEN VULNERABLE ROAD USERS AND AUTONOMOUS VEHICLES

New driverless vehicles are about to be introduced on the roads. In some cities, public authorities are conducting tests to assess whether the circulation of autonomous vehicles (AVs) will lead to operational and safety issues. One of the main concerns regard the interaction between AVs and pedestrians, which requires attention and ability for pedestrians to interact correctly with driverless vehicles, and, on the other side, for driverless vehicles to detect pedestrians thus avoiding collisions. One common approach to studying safety conditions consists in carrying out a conflict analysis. Conflict is reached when the entities are close to collision. The latter is avoided by the execution of an evasive manoeuvre by at least one of the two conflicting moving entities. This thesis deals with the activities conducted by the city of Turin as part of the European SHOW project. In September and October 2022, two Navya autonomous shuttles had circulated in the Lingotto district. Field observations were conducted at an un-signalized pedestrian crosswalk, where the interactions between pedestrians, AV and ordinary vehicle were captured though video recordings. A camera was placed on top of a telescopic pole and recorded a total of 24 hours over 8 days. Videos were first corrected to eliminate distortion, then the georeferenced spatial temporal trajectories of pedestrians and vehicles were extracted through a proprietary C++ code. They were used to estimate kinematic variables after removing noise in data using a Simple Moving Average filter. A proprietary Python code was developed to evaluate surrogate safety measures (SSM) like Time-to-Collision (TTC) in the pre-event phase, and Post Encroachment Time (PET) in the post-event one, to get a quantitative estimation of observed conflicts. Other secondary SSMs were derived from the first two. Statistical tests were conducted on 33 AV- and 135 traditional vehicle pedestrian interactions. For both TTC and PET values, the log-normal probability distribution resulted to be the most suitable. The analysis on TTC data revealed that the two distributions were not significantly different. However, the two distributions showed clear dissimilarities along the tails. In particular, differences were evident in the lower tail of TTC probability distribution function, where data from pedestrian-ordinary vehicle interaction showed lower TTC, i.e. worst values than AV pedestrian ones. Therefore, the probability of a dangerous conflict is higher in the case of pedestrian-human driven vehicle conflict. The analysis on PET data reveals that the two distributions were significantly different. The pedestrian-AV distribution showed larger PET values, which reflects the increased caution taken in scheduling driverless vehicles that, after a stop, restart cautiously to repeatedly check for other potential conflicts with moving entities in the surrounding. Results on PET also explain the way that AVs operate in a traffic network where the human factor is still predominant. Now, AVs operate at low speed and stop whenever a moving obstacle (e.g., a pedestrian) is detected. Although they can negatively influence both capacity and level of service of the road network, AV operations are safe as the outcomes of this study reveal.