A 3D target detection algorithm for low-speed unmanned vehicles in closed parks based on redundant fusion of multi-sensor information

To address the issues of object segmentation and pedestrian mis-detection, this paper proposes a redundant target detection method based on multi-level Euclidean clustering and view cone point cloud fusion. First, the joint calibration of LiDAR and camera is completed. Then, the pedestrian information and point cloud fusion are used to form a pedestrian cone point cloud, in addition, the multi-level threshold Euclidean clustering algorithm and the optimal 3D bracket selection method are designed. Finally, the pedestrian 3D bounding box is obtained by solving the point cloud confidence function, and fusion matching with the LiDAR detection results is performed to output the multi-sensor fusion perception results. Real-vehicle experimental data show that this method improves the accuracy of the whole sensing module, reduces the number of missed and misdetected boxes, and achieves 94.94% detection accuracy, which is 8.68% higher than the LiDAR detection algorithm, demonstrating its effectiveness and reliability.