A Residual Graph Networks for 3D object detection in point clouds
Wenju Li, Zhilin Chen, Qu Jiantao, Liu Cui · 2023
The text describes a 3D point cloud object detection model based on a graph neural network. Due to the disorder, sparsity and overlap of LiDAR point cloud, it leads to problems in the detection accuracy of small objects. To solve this problem, a 3D small object detection algorithm based on residual graph neural network is proposed. The algorithm improves performance through two key steps. First, the original point cloud is sampled, and the sampled point cloud is framed within a fixed radius. Second, residual connections are introduced to enhance the feature extraction capability of graph networks. Finally, the aggregation of node features is completed on the deep feature map, and the features of the two stages are spliced. Experimental results show that on the KITTI dataset, the proposed method achieves significant performance improvements in point cloud small object detection tasks with higher accuracy and discrimination. By employing residual connections, the model can better capture important features in point clouds, thus improving the feature extraction capability of graph networks.