High performance point-Voxel feature set abstraction with mamba for 3D object detection

In the field of autonomous driving, a two-stage three-dimensional object detection approach has seen significant advancements. However, challenges persist in terms of detection accuracy, which can have a profound impact on the safety of autonomous vehicles. This study examined four critical issues that impair the accuracy and efficiency of the model: limited acceptance fields, slow acquisition of global features from voxels, challenges in capturing keypoint features, and uncertainties associated with network post-processing. To address these challenges, we propose four novel techniques: (1) a non-empty voxel feature extraction method that utilises linear angular attention to broaden the receptive field; (2) an efficient voxel feature extraction and downsampling approach based on Mamba2, designed to accelerate the acquisition of global voxel features; (3) a node extraction strategy that employs the Kolmogorov-Arnold Network (KAN) to extract key point features via segmented farthest point sampling (S-FPS); (4) a fuzzy non-maximum suppression (Fuzzy-NMS) method that refines suppression thresholds during the post-processing phase. By integrating these techniques, we introduce a High-Performance Point-Voxel Region Convolutional Neural Network (HP-PV-RCNN) algorithm specifically tailored for precise 3D object detection. We validated the effectiveness of the HP-PV-RCNN algorithm through comprehensive experiments using the Kitti, NuScenes, and Waymo open datasets. Specifically, our proposed network attained average precisions of 83.73 % for vehicles, 76.32 % for bicycles, and 53.52 % for pedestrians in the medium-difficulty category of the Kitti dataset for detecting these entities. The code and model are available at https://github.com/jlauwcj/HP-PV-RCNN.