Probabilistic motion prediction of surrounding vehicles via artificial neural network and its application for longitudinal collision avoidance system

Abstract

Since global companies such as Google had recently commercialized the autonomous driving vehicles, research towards the autonomous driving has been actively conducted over the last decade. However, the technology of handling fatalities in risky driving situations is still insufficient and most customers are concerned about the safety of autonomous vehicles. To resolve this safety problem, autonomous vehicles are required to detect potential dangers in future situations and mitigate accidents in advance. In addition, the future uncertainty should be considered because a deterministic single output cannot represent uncertain future properly. Therefore, it is essential to have a reliable probabilistic prediction algorithm which can foresee vehicle motions in the future and handle future uncertainties. This thesis proposes the probabilistic motion prediction algorithm that can accurately computes the probabilities of multiple target lanes and trajectories of surrounding vehicles by using the artificial neural network

more specifically radial base function network (RBFN). RBFN prediction algorithm can estimate the future uncertainty in the categorical distribution of which lane the vehicle is going to arrive in near future in a probabilistic manner and can represent multiple future trajectories in parallel in order to handle the uncertainties of future position. In order to demonstrate and verify the effectiveness of the proposed prediction algorithm, it is applied for the longitudinal collision avoidance control problem. Since the proposed RBFN prediction algorithm represents future uncertainties, chance constrained model predictive control (CCMPC) is utilized because the collision chance constraints in CCMPC deal with the uncertainty of collision based on the future uncertainty. The RBFN-based CCMPC simulation is conducted for several risky cut-in scenarios and compared with the Interactive Multiple Model (IMM)-based prediction algorithm. Simulation results show that the RBFN-based CCMPC uses smaller inputs for collision avoidance but achieves higher comforts when compared to the IMM-based CCMPC. In addition, owing to the 1.1 seconds earlier predictions, RBFN-based CCMPC achieves higher collision success rate than the IMM-based CCMPC, and thus can guarantee higher safety margins for all levels of defensive driving