3D Detection for Occluded Vehicles From Point Clouds

Abstract

In autonomous driving, 3D vehicle detection plays an important role in traffic safety, and the detection of occluded vehicles is equally important. Different from 2D images, the point clouds are sparse and disordered, causing an uncompetitive expression for occluded vehicles. As a result, current detection methods cannot achieve a higher performance for the detection of occluded vehicles. To alleviate this problem, we present a novel 3D detector for occluded vehicle (3DOV) detection using lidar data. The proposed method has the following contributions: different data representations, including a bird’s-eye view and voxel grids, which are fused to strongly represent the features of occluded vehicles; a novel voxel-wise convolutional encoder that leverages the hierarchical property of convolutional networks to learn features with spatially local correlation from unordered points; a multiscale detection header that removes the influence of occlusions and optimizes the backpropagation of gradients; and a slimmer 3DOV network is presented to handle the speed-accuracy tradeoff. The experiment results on the KITTI 3D vehicle-detection benchmark show that the 3DOV has achieved excellent performance in terms of accuracy and efficiency, especially for occluded vehicles. Ablation studies are provided to demonstrate the effectiveness of each designed module.