3D Object Detection Based on Neighborhood Graph Search in Dense Scenes

Abstract

In the field of robotics and autonomous driving, achieving accurate 3D object detection is crucial for the perception of complex traffic environments. Most current research uses deep learning methods to extract object features from point clouds or images. However, these approaches often do not fully utilize the mutual positional information between objects, resulting in low detection accuracy in dense scenes. To address this issue, this paper proposes a frustum point cloud 3D target detection algorithm based on the fusion of camera and LiDAR data. We establish global connectivity of objects based on the degree of overlap between camera detection frames. Then, we design a neighborhood graph search algorithm based on constraint satisfaction to match the camera target detection results with LiDAR clustering results. Finally, the category and distance information of obstacles are displayed in bird's eye view (BEV). Evaluated on the KITTI benchmarks, our method achieves an average precision (AP) of 88.38% on the easy level, 87.64% on the moderate level, and 80.10% on the hard level in BEV detection. Compared to F-ConvNet, our method shows improvements of 5.49% on the moderate level and 7.33% on the hard level, significantly enhancing recognition accuracy in dense scenes.