3D Object Detection Based on Feature Fusion of Point Cloud Sequences

Increasing the number of lines of lidar can alleviate the sparsity of point cloud, but the cost of high-line lidar is higher. From the perspective of algorithm, the fusion of continuous point clouds frames can obtain more abundant information, which is expected to be a means to achieve high-performance object detection on low-line lidar. At present, most multi-frame point cloud fusion methods stack different frames after registration. This fusion method, which introduces GPS, ego-motion, and other information for registration, can only align static objects, not larger moving objects. In this paper, we implement a multi-frame fusion method based on Non-Local, which does not need registration and takes into account both moving and stationary objects. We validated the model performance on the NuScenes dataset. Experimental results show that the performance of the proposed fusion method is better than that of fusion method by stacking frames after registration.