Position-Aware Voxel Aggregate Network for Two-Stage 3-D Object Detector

Voxel-based 3-D object detection methods have drawn extensive attention due to their efficiency and accuracy. However, effectively using the feature information of point clouds to improve the detection performance remains to be explored. In this article, we propose a novel two-stage object detector, called position-aware voxel aggregate network for two-stage 3-D object detector (PA-Det3D). First, in the region proposal network (RPN) stage, we introduce a cross-scale location attention module (CLM) that enables effective fusion of high-level semantic information and low-level spatial details at each location. Second, we introduce a secondary grouping module (SGM), which consists of two components: the position offset secondary grouping module (PSM) and the distance-enhanced module (DEM). The PSM predicts the relative positional offsets of querying voxels, allowing for dynamic adjustment of query positions to overcome distance limitations. The DEM, on the other hand, incorporates relative distance information to provide more explicit geometric properties and enhance robustness. Finally, we propose a consistency matching based on the Gaussian mixture model (CMG), which can adaptively adjust the threshold to establish teacher–student relationships, resulting in more stable supervision signals. Extensive experiment results on the KITTI and Waymo datasets demonstrate that our approach achieves competitive performance in 3-D bounding box detection, which verifies the effectiveness of our method.