Caracterización del comportamiento de los usuarios vulnerables de la vía mediante técnicas de Realidad Virtual y modelos de Machine Learning: Characterization of vulnerable road user behavior using Virtual Reality techniques and Machine Learning models

Vulnerable Road Users (VRU) include pedestrians, cyclists, motorcyclists, and electric scooter users. In Spain and in the European Union, pedestrian and cyclist accidents and injury rates have been growing steadily in recent years, especially in urban environments.

Therefore, this doctoral research has focused on the implementation of two methods to characterize the behavior of pedestrians and cyclists in potential collision situations in the city: Virtual Reality (VR), and video surveillance and camera recording systems. Additionally, and considering the extent and level of detail of the pedestrian accident database compared to the cyclist counterpart, a technological application is developed based on the integration of a predictive pedestrian collision model to optimize the Advanced Driver-Assistance Systems (ADAS) in a commercial vehicle.

The research activities, carried out at the Instituto Universitario de Investigación del Automóvil Francisco Aparicio Izquierdo (INSIA), have been funded by the OPREVU, VULNEUREA, SAFEDUCA, VIRESTREEP projects and partially by the SEGVAUTO program. The virtual environments were designed by the Centro de Domótica Integral de la UPM (CEDINT), and are based on accident data in Madrid and an accident report from the Netherlands Organization for Applied Scientific Research (TNO).

In the experimental session, a portable backpack equipment coupled to a VR helmet has been used. Also, the bicycle simulator includes a controller, a training roller and an ad-hoc platform with pulley system to synchronize turning, speed and braking. In the simulations with cyclists, eye-tracking functionality has also been integrated for vehicle identification.

From the VR results, predictive collision models have been generated by supervised Machine Learning classification, based on the VRU's kinematics, attentional level and visual perception.

As part of the study using camera surveillance and recording systems, traffic flows in Vienna (Austria) have been analyzed using the Mobility Observation Box (MOB) system during the doctoral stay at AIT Austrian Institute of Technology GmbH. The obtained database of interactions was classified by clustering and validated by supervised Machine Learning classification in order to identify patterns of user behavior and establish up to three levels of road risk.

In the technological innovation section, an Autonomous Emergency Braking (AEB) system of a commercial vehicle was evaluated on track through validation tests, with the aim of modeling its decision algorithm. Its possible optimization areas were identified by subsequent simulation in CarSim, based on VR behavior patterns.

Given that one of the most influential variables in the pedestrian collision predictive model is the percentage of time that pedestrians observe the vehicle approach zone, a facial and eye recognition system has been designed that can be integrated into the AEB system, capable of determining whether or not the pedestrian is observing the vehicle and his or her level of attention.

In the optimization phase, a prototype has also been designed and simulated in Simulink and CarSim that integrates the optimized AEB system with the predictive pedestrian collision model, together with an Automatic Evasive Steering System (AES) and a series of additional ADAS. The evaluation of the joint system (OPREVU-AES) includes the analysis of the reconstructed crash avoidance rate and the variation of the probability of severe head injury (ISP) using PCCrash.

As conclusions, the models resulting from the VR and MOB system tests share a number of common kinematic variables that characterize the behavior of VRUs in potential collision situations. The accuracies of the generated models exceed 83%, with a great balance in the performance metrics. For its part, the face and eye recognition system guarantees an overall detection accuracy of over 79%.

Finally, OPREVU-AES can generate stable trajectories between 40 and 70 km/h, increase the available braking distance for a following vehicle and the reaction time of its driver, reduce the probability of serious head injury by 65% and an avoidance rate of up to 77.9%.