Anchor-Free Spatio-Temporal Attention Pyramid Network for Object Detection in Driving Scenes

In driving scenarios, perception and motion prediction are important guarantees for the safe and stable operation of autonomous vehicles. In recent years, an increasing number of studies have directly used raw point cloud data as input for perception and motion prediction. In this paper, we propose a new anchor-free spatio -temporal attention pyramid network that enhances the network's perception ability of driving scene objects by obtaining multiscale spatio -temporal contextual information, while substantially improving the accuracy of object detection in the network on the basis of ensuring the effectiveness of motion prediction. Experimental results on the nuScenes dataset show that our method significantly outperforms all state-of-the-art LiDAR-based networks in terms of perception.