Spatio-Temporal Fusion: A Fusion Approach for Point Cloud Sparsity Problem

3D LiDAR point cloud target detection is one of the keys of robot and autonomous vehicle. To low-beam LiDAR, missing detection is inevitable because of sparsity of the point cloud. To solve the problem, a fusion method based on temporal and spatial dimensional information is proposed to increase the richness of the point cloud to improve the accuracy of target detection. The method consists of two stages. Firstly, in order to solve the uncertainty of point cloud data density cause by Down-sampling, local and global features of the point cloud data in single frame are fused. Point cloud density is estimated to obtain local feature of the point cloud, then a combined global and local point cloud feature enhancement method is used for feature fusion, and the feature-enhanced single-frame point clouds are acquired by bilateral filtering. Secondly, to solve the problem of low accuracy of target detection caused by irrelevant data, the correlation coefficients of multi-frame point cloud data from last stage are evaluated by Hausdorff distance, and the point cloud data with low correlation degrees according to the correlation coefficients are eliminated to achieve the correlation fusion of multi-frame point clouds. Finally, the point cloud data with richer features is obtained to improve the target detection accuracy of 3D point clouds. The experimental results from NuScenes dataset and real outdoor scenes show the effectiveness of proposed method.