A graph attention feature pyramid network for 3D object detection in point clouds

Graph Neural Networks have been recently applied to 3D object detection in point clouds. The works, however, have the problem of insufficient detection accuracy for small objects and objects in complex backgrounds. Towards this end, a graph attention feature pyramid network is proposed for 3D point clouds object detection. Specifically, the network constructs a near-neighbors graph in a point cloud which is downsampled; and then, a graph attention feature pyramid network is designed to extract features of the point cloud at different levels; finally, a feature fusion module is employed to fuse the features before point classification and object detection. Compared with the benchmark network Point-GNN on the KITTI dataset, the detection accuracy of cars in complex scenes is improved by 2.53%, and the detection accuracy of pedestrian and cyclist categories in moderate scenes and complex scenes is improved by 5.17%. and 3.62%. The experiments show that the designed method is more effective for the detection of small objects and objects in complex backgrounds in 3D object detection.