NSAW: An Efficient and Accurate Transformer for Vehicle LiDAR Object Detection

Nowadays, light detection and ranging (LiDAR) plays an important role in environmental perception technology; however, LiDAR cannot detect some blocked objects, and it can even detect some incorrect and unimportant objects as target objects due to the sparsity and disorder of the point cloud and the limited receiving area. Indeed, the detection accuracy can be improved by setting multiple cameras and LiDAR, but the cost can be unacceptable. In order to resolve these issues, a transformer-based architecture that can connect voxels with each other by neighborhood sparse attention is proposed. A standardized transformer, however, cannot be applied directly to sparse point cloud space because it contains a large number of empty voxels. Hence, a scheme of neighborhood sparse voxel attention to window attention (NSAW) is introduced, which is an efficient and accurate single stride transformer for connecting 3-D backbone to 2-D bird-eye-view (BEV) feature map. In this architecture, neighborhood window attention (NWA) and neighborhood voxel attention (NVA) are proposed to effectively extract features from no-empty voxels, which can achieve 84.30% average precision with 40 recall points (AP3D(R40)) and bring 2.68% AP3D(R40) improvement for Voxel region-based convolutional neural network (R-CNN) on the moderate car class on the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) test split. On the premise of computational efficiency, the proposed NSAW demonstrates quite a few improvements over the traditional baselines on KITTI3D dataset.