Spatial-Temporal Attentive Motion Planning Network for Autonomous Vehicles

Abstract

Different deep learning approaches have been devised for Autonomous Vehicle (AV) motion planning. However, most of such learning approaches rely on generic visual features. To efficiently generate feasible trajectories, we introduce Spatial-Temporal Attentive Motion Planning Network (STAMPNet). STAMPNet consists of AttentiveFC and Attentive 3D-CNN modules. The former encodes the trajectory histories, and the latter models the Spatial-temporal representations from videos. Then we fuse both attentive intermediate features and feed them into an LSTM that outputs planned trajectories. Experimental results demonstrate that SAMPNet effectively generates feasible trajectories for autonomous vehicles.