A reinforcement learning intelligent deductive model with pre-trained sequence information

Abstract

Agent trajectory prediction is an increasingly popular topic in computer vision and autonomous driving. With the help of deep learning and big data, it is possible to understand the interaction model between agents hidden in complex environments. Existing methods usually pay more attention to the average trajectory offset of the agent while ignoring the distribution differences of the target. This issue results inevitable performance decrease. To address this issue, we propose a novel reinforcement learning intelligent deduction model (RLDM). It achieves joint reasoning of goals and paths in a unified framework, and accurately predicts trajectories in a short period of time with fewer datasets. Specifically, an end-to-end time-series pre-training module is proposed to explore the agent's training state reward and goal reward. Moreover, a prediction module based on the combination of kinematics and environmental background is proposed to explore the agent motion characteristics. By this way, acting in a purely reactive manner is better relieved. Practical trajectory prediction experiments are designed, and the experimental results verify the superior performance of our proposed model. The model experiment results are improved by 2% and 11% on the ADE and FDE metric on average.