Global 3D Point Cloud Object Detection System Based on Data-Level Stitching

With the rapid development of artificial intelligence, the application prospects of global perception systems that can cover large-scale smart tunnel scenes are becoming increasingly widespread. Using multi-modal data from different sensors, the global perception system attempts to locate and track traffic targets. Due to the presence of detection blind spots at a considerable distance between two stations, which increases the difficulty of detection, the conventional global stitching method based on result-level stitching easily leads to problems such as lost vehicles and discontinuous trajectories in the blind area, and the low detection accuracy of sparse point cloud detection at the single station. To address these issues, this paper optimised the point cloud detection algorithm by improving the network structure and loss function to enhance the perception capability of the single station. Additionally, it proposed a data-level global point cloud stitching algorithm and a method for sampling from a difficult database, replacing the traditional global result-level stitching method and ensuring the fusion effect of global trajectories. Overall, this provides a more reliable and comprehensive perception fusion result for platform twinning. Finally, to validate the effectiveness of our method, we introduced the publicly available VANJEE-PointCloud dataset collected in the real world. The experiments show that our algorithm not only enhances perception capability but also improves the success rate of global trajectory fusion.