Continuous frames of point-cloud-based object detection is a new research direction. Currently, most research studies fuse multi-frame point clouds using concatenation-based methods. The method aligns different frames by using information on GPS, IMU, etc. However, this fusion method can only align static objects and not moving objects. In this paper, we proposed a non-local-based multi-scale feature fusion method, which can handle both moving and static objects without GPS- and IMU-based registrations. Considering that non-local methods are resource-consuming, we proposed a novel simplified non-local block based on the sparsity of the point cloud. By filtering out empty units, memory consumption decreased by 99.93%. In addition, triple attention is adopted to enhance the key information on the object and suppresses background noise, further benefiting non-local-based feature fusion methods. Finally, we verify the method based on PointPillars and CenterPoint. Experimental results show that the mAP of the proposed method improved by 3.9% and 4.1% in mAP compared with concatenation-based fusion modules, PointPillars-2 and CenterPoint-2, respectively. In addition, the proposed network outperforms powerful 3D-VID by 1.2% in mAP.
Keywords: 3D object detection; attention mechanism; autonomous driving; feature fusion; point cloud sequences.