A Small-Size 3d Object Detection Network for Analyzing the Sparsity of Raw Lidar Point Cloud

Abstract

LiDAR is one of the main sensors for 3D object detection in autonomous driving. LiDAR has the advantages of high precision and high resolution, but as the distance increases, the points it acquires become sparse, resulting in uneven sampling points and hindering the feature extraction of discrete objects. Current 3D object detection methods using LiDAR ignore the sparse features of the original LiDAR point cloud, resulting in low classification accuracy over small object detection, hindering the development of autonomous driving technology. To address this problem, we propose point cloud Sparse Detection Network (PCSD), an end-to-end two-stage 3D object detection framework. First, PCSD uses a data augmentation algorithm to preprocess the KITTI dataset, then uses voxel point centroids to locate voxel features, and then uses a point sparsity-aware RoI grid pooling module to aggregate local voxel features. Finally, we improve the confidence of the final bounding box by using voxel features with the original point cloud sparse features. Experimental evaluation on the challenging KITTI object detection benchmark shows significant improvements, especially in pedestrian and cyclist classification accuracy improved by 13.22% and 9.33%, respectively, demonstrating the feasibility and applicability of our work.