Dynamic obstacle avoidance strategy for autonomous vehicles based on LSTM model trajectory prediction under urban roads

Abstract

Studying obstacle avoidance strategies for self-driving cars is important research direction. We propose a LSTM-based model for predicting the trajectories of “obstacles” around the self-driving vehicle that have an impact on its motion state, so that the self-driving vehicle can carry out less risky path planning in advance to improve the safety and stability of the self-driving vehicle. First, the LSTM network is trained using the NGSIM dataset to obtain a model that can more accurately predict obstacle motion information. Then, the obstacle risk function is used as a sequential judgment index for obstacle trajectory prediction by constructing a prediction level range model as an obstacle predictable region. According to the obstacle prediction information to derive its movement trajectory in a period of prediction time domain, the self-driving car can predict the movement trajectory of the obstacle in advance, and the planning layer plans the path that can avoid the obstacle, to ensure that the vehicle travels stably and safely. Finally, the predicted trajectory is combined with the planning layer to ensure safe obstacle avoidance by the vehicle using multi-constraint OCF-MPC controller. Experiments show that the use of trajectory prediction scheme can reduce the risk of self-driving vehicle traveling.