CRobot-SECOND: an improved SECOND-based 3D point cloud object detection method for safety inspection robots in building construction sites

Abstract In safety inspection tasks in building construction sites, 3D point cloud object detection is a key technology for safety inspection robots to perceive construction workers, construction vehicles, and surrounding obstacles. However, sparse low-channel Light Detection and Ranging (LiDAR) point clouds, dynamic-object interference, and deployment on resource-constrained robotic platforms make it difficult to improve detection accuracy without compromising detector-level inference efficiency. To address this problem, this study proposes CRobot-SECOND, a 3D point cloud object detection method that combines point cloud input enhancement with lightweight backbone optimization. First, a point cloud fusion method based on ground constraints and dynamic object compensation is constructed to improve point cloud density and geometric completeness while suppressing ghosting artifacts. Then, the Partial Convolution (PConv) module and the Simple Parameter-Free Attention Module (SimAM) are introduced into the two-dimensional Bird’s-Eye-View (BEV) backbone of Sparsely Embedded Convolutional Detection (SECOND) to reduce redundant computation and enhance key-region feature representation. Finally, a construction-site point cloud dataset containing construction workers and construction vehicles is established, and comparative, ablation, and multi-seed experiments are conducted. The experimental results show that the proposed fusion method improves point cloud density while maintaining high fusion quality. Across five independent training seeds, CRobot-SECOND achieves 3D average precision (AP) values of (57.77 ± 0.50)% for construction workers and (82.84 ± 0.37)% for construction vehicles, compared with (53.09 ± 0.44)% and (81.08 ± 0.32)% for SECOND, respectively, where the results are reported as mean ± standard deviation. The category-averaged 3D AP increases from (67.08 ± 0.35)% to (70.31 ± 0.43)%, with a mean paired improvement of 3.22 ± 0.10 percentage points. The detector-only network inference speed increases from 75.19 Hz to 92.08 Hz. Overall, these results show that coordinated point cloud input enhancement and lightweight detector optimization improve frame-level 3D detection of construction workers and construction vehicles under the collected construction-site conditions without increasing model scale. These frame-level detection results provide a basis for exploring CRobot-SECOND as a perception component for construction-site safety monitoring, robot path planning, and obstacle avoidance.