An Integrated Framework of Autonomous Vehicles Based on Distributed Potential Field in BEV.

This paper presents an integrated framework of decision-making and motion planning for autonomous vehicles (AV) based on real-time trajectory prediction and distributed potential field in Bird's-Eye View (BEV). Firstly, the distributed tracking system is applied for global trajectory tracking of traffic participants in BEV, which can avoid the physical limitations of onboard sensors. Then, the end-to-end network model BiLSTM-MDN is utilized for the trajectory prediction of traffic participants. To cope with the prediction error caused by other neural networks outputting only one prediction trajectory, the model BiLSTM-MDN outputs a probability density function (PDF) generated by Gaussian Mixture Model (GMM). At the same time, the PDF is adopted in the potential field model, which describes road constraints and traffic participants with diverse behaviors by different potential functions. In addition, the Frenet Coordinates substitute for the Cartesian Coordinates to simplify the calculation. Eventually, a constrained multi-objective optimization problem is proposed to integrate decision-making and motion planning. It is reduced to a P problem and solved by the Depth-First-Search (DFS) algorithm to obtain an optimal solution. The typical intersection scenario is taken to validate the performance of the presented approach. Simulation results show that the integrated system is feasible and effective because it can address various actions of other traffic participants adequately through appropriate decisions and robust planning.