3D object detection algorithm based on adaptive geometry-weighted density perception

Abstract In 3D object detection for autonomous driving, object scale and occlusion exacerbate object sparsity, thereby reducing detection accuracy. To address this issue, this paper proposes a 3D object detection method based on adaptive geometric density weighting. First, a CADR module is designed. This module dynamically perceives spatial geometric structures by calculating the curvature of local point clouds and adaptively assigns weights based on neighborhood density. It assigns higher semantic weights to keypoints in sparse regions, thereby effectively enhancing the features of low-density objects. Additionally, a Point-level Multi-attention Module is constructed. This module utilizes multi-branch pooling and adaptive fusion to extract local features, effectively suppressing background interference and outlier noise. Experiments were conducted on the Waymo-mini and KITTI datasets, and the method was compared with other 3D object detection algorithms. Experimental results show that, in 3D mode, the proposed algorithm achieves a mean Average Precision of 72.78%. Overall, the proposed method significantly enhances sparse point cloud features, further improving the accuracy and stability of 3D object detection in complex scenes, and provides a reliable technical solution for autonomous driving environmental perception systems.