Feature Pyramid Network Based on Skip Connect for Vehicle Prediction

Abstract

The trajectory prediction of the target vehicle can improve the perception ability of autonomous driving vehicles in the open world environment and is an essential technology for realizing high-level fully automatic driving. The traditional vehicle trajectory prediction model based on image input is sensitive to the image style of the dataset and has weak generalization ability. The prediction model based on intermediate features can overcome the problem of different dataset styles. However, there are relatively few explorations on building models based on intermediate semantic features. Inspired by the feature pyramid in image processing, this paper proposes an FPSP (Feature Pyramid network based on Skip connect for vehicle Prediction) network. By adding the pyramid feature branch, the original network is improved. The initial coarse information is hierarchically introduced into corresponding abstract high-dimensional features. Experiments on the KITTI dataset prove that FPSP achieves considerable improvement with only a slight increase in time consumption.