Vehicle Instance Segmentation and Prediction in Bird's-Eye-View for Autonomous Driving

Joint perception and prediction methods predict the segmentation of traffic participant instances and their future movements under the BEV spatial representation, using raw sen-sor inputs in an end-to-end manner. However, existing research still faces the problems of large model volume and slow inference speed, which cannot meet the requirements of lightweight and fast autonomous driving systems in practical applications. To ad-dress this issue, this paper implements an efficient and lightweight end-to-end joint perception and prediction model. The model is built based on CNN only and is implemented through two stages of training. The first stage builds a BEV semantic segmentation model to obtain features of the scene of BEV space and various types of traffic participants. The second stage predicts vehicle instance segmentation and future movements based on the BEV features trained in the first stage. Experiments on the nuScenes dataset show that the model in this paper improves the inference speed to approximately two times that of the existing studies and shrinks the number of model parameters to 1/2 without losing too much prediction performance. The ablation experiments validate the effectiveness of each module on the overall performance.