Research on target recognition algorithm of ground point cloud on push-broom LiDAR scanning

To address the challenge of target detection in complex backgrounds using push-broom LiDAR scanning ground point clouds, this study proposes an improved PointNet⁺⁺ network based on a self-attention mechanism, enhancing the algorithm's ability to capture local neighborhood features. The algorithm's capability for complex point cloud scene segmentation was validated using the semanticKITTI dataset. Subsequently, to precisely locate the key points of the target, the point cloud segmentation results were utilized to extract the key parts of the target point cloud, and the geometric center was calculated. A laser dynamic scanning ground point cloud target detection test system was established to acquire three-dimensional point cloud datasets of typical ground targets, and the algorithm was validated. Compared to PointNet⁺⁺ and RandLA-Net, the improved PointNet⁺⁺ network with the self-attention mechanism exhibits enhanced recognition capability for smaller ground targets and greater generalization ability across various complex scenarios. This study is the first to combine point cloud scene segmentation with target key point localization, verifying the applicability of the improved PointNet⁺⁺ network in battlefield environments and achieving innovation at the application level.