Abstract:Point cloud object detection technology can provide crucial support for fields such as autonomous driving and robot environmental perception. However, point cloud data typically exhibit unstructured characteristics, uneven spatial distribution, and susceptibility to noise interference, which pose significant challenges for traditional methods to achieve complete segmentation and precise object extraction, complicating subsequent data processing tasks. To tackle issues of under-segmentation and over-segmentation common in traditional Euclidean clustering algorithms, this paper proposes a smoothness-based Euclidean clustering method. By introducing a smoothness threshold, the proposed approach effectively identifies edge regions of targets in point clouds, thereby enhancing clustering accuracy. By introducing principal component analysis, background and noise point clouds are effectively eliminated to retain key targets, thereby enhancing the efficiency of feature matching and recognition. Experimental validation on the KITTI public dataset, simulated sea-surface dataset, and Geiger-mode avalanche photodiode photon-counting radar measured data demonstrates that the proposed method effectively mitigates the problems of over-segmentation and under-segmentation, significantly improving the accuracy and robustness of point cloud object detection.