SAPT: A Hierarchical Sliced Anchor Point Transformer for LiDAR Point Cloud in Robot Perception

ABSTRACT In robot perception, convolutional networks are representative of hierarchical neural networks and have achieved outstanding results in fields of text and image processing. The key to extracting local features and integrating them into global features is the presence of shared regions between windows, which is a significant factor in the success of convolutional neural networks. Inspired by this success, this paper investigates existing methods on hierarchical point cloud network methods for grouping point clouds and proposes a solution involving shared anchor points to enable subsets to share points. This effectively enhances the feature extraction capability of existing hierarchical point cloud networks. Building upon this, we have designed a slice‐based anchor point transformer called SAPT, which exhibits strong compatibility with downstream tasks. Experimental results demonstrate that, with assistance of shared anchor points, extracted features are not only suitable for classification and segmentation tasks but also perform exceptionally well in object detection tasks, without the need for designing specific network enhancements for local feature extraction. Empirical evidence shows that SAPT excels in both classification and segmentation as well as object detection tasks. On KITTI (LiDAR only), with a threshold of 0.7 IoU and 40 recall points, the mAP value reached a maximum of 79.85.