Probabilistic Collision Estimation System for Autonomous Vehicles : Evaluated in Intersection Scenarios Using a Velocity Planning Controller

Abstract: Nearly 1.3 million people die each year in traffic-related accidents, whereas an additional 20-50 million people are injured. Introducing autonomous vehicles into traffic would aim at reducing accidents and improving the transportation experience for humans. However, intersections are still difficult traffic scenarios for such systems due to their complex nature. Functions which could accurately foresee future events would improve autonomous systems when dealing with such complex traffic scenarios. This thesis is carried out at ÅF AB, Stockholm, in collaboration with the Integrated Transport and Research Labs at KTH. The aim is to research how a collision prediction system could be implemented on the concept vehicle developed by the Transport Labs. This system would give the vehicle the ability to make decisions of estimated future events based on information gathered in real-time. A conceptual collision prediction system is developed using the Robotics Operating System simulation environment. The simulations are conducted on T-intersection traffic scenarios where only the ego vehicle and another observed vehicleis present. The system has two outputs, 1) probability fields which indicates probable future positions of the observed car and 2) a summarized collision probability which shows how probable collisions are over the prediction horizon given a certain control input for the ego vehicle. The system performance is evaluated using a model predictive velocity controller which chooses the control input that gives the lowest value for the summarized collision probability. In the current state, both parts are too slow to work in real-time applications, but this is something that is achievable with the correct implementation. The results for system robustness and precision are promising and the vehicle successfully avoids collisions given that the sensor noise is low enough and it is shown that this is a promising system.