Background Filtering and Target Detection for Traffic Objects with Roadside LiDAR Point Cloud Data

Roadside perception technology is an essential component of traffic perception technology, primarily relying on various high-performance sensors. Among these, LiDAR stands out as one of the most effective sensors due to its high precision and wide detection range, offering extensive application prospects. This study proposes a voxel density-nearest neighbor background filtering method for roadside LiDAR point cloud data. Firstly, based on the relatively fixed nature of roadside background point clouds, a point cloud filtering method combining voxel density and nearest neighbor is proposed. This method involves voxelizing the point cloud data and using voxel grid density to filter background point clouds, then the results are processed through a neighbor point frame sequence to calculate the average distance of the specified points and compare with a distance threshold to complete accurate background filtering. Secondly, a VGG16-Pointpillars model is proposed, incorporating a CNN network during the point cloud encoding process and adding average pooling weights to enhance point cloud features. The backbone network uses the VGG16 network to extract feature maps of different scales and adds concatenation layers to improve detection accuracy. This method can filter out 99.74% of background point clouds and improve the mean average precision of the target detection model by 3.04%. The model's practical applicability is demonstrated through transfer applications on real-time data.