DENSIM3D: Enhancing 3D Object Detection through Density Weighting and Similarity Analysis
Yue Sun, Ke Lü, Shuhua Wang, Jian Xue · 2023
LIDAR has become one of the primary sensors for 3D perception in autonomous driving technology. The sparse and discrete nature of the point cloud generated by LIDAR presents significant challenges for 3D object detection. In this paper, we propose a lightweight network based on point cloud, which offers high efficiency and accuracy for 3D object detection. To mitigate the negative impact of point cloud sparsity on detection accuracy, we propose a density-weighted feature extraction structure called the Density Weighting Module (DWM). This structure estimates the density of the point cloud and improves the feature extraction process through optimized weighted means. In detection scenarios, there is a notable difference in accuracy when detecting objects of different sizes, particularly with lower accuracy for small objects due to the limited number of positive samples. To mitigate the impact of this problem, we attempt to differentiate the detection points of objects of various sizes through similarity analysis. We develop the Similarity Weighting Module (SWM) to enhance the representation capabilities of small objects by computing similarity, thereby highlighting their distinctive features. To improve detection accuracy while reducing computational costs, we propose a fusion strategy called Foreground-Background Fusion with Channel Attention (FBA). This strategy combines information from the foreground and background points and incorporates channel attention mechanisms to avoid information redundancy. Through experimental evaluations on the widely used KITTI dataset, our method outperforms most state-of-the-art single-stage methods of point cloud-based object detection.