Speakers
Description
Recent years have witnessed the increasing deployment of construction robots in construction sites. To improve the collaboration efficiency between workers and construction robots, as well as to reduce the occurrence of accidents, 3D-based detailed perception methods have become more essential, especially for human body part detection. In this regard, this study proposes a LiDAR-based deep learning model for the 3D detection and pose estimation of workers on construction sites to predict the workers’ 3D bounding boxes and body key-points. To enhance prediction accuracy under various poses, the network employs a local attention mechanism combined with 3D sparse convolution to capture detailed information from local sparse voxels, while a multi-scale fusion module is utilized to integrate point cloud features at different scales, thereby obtaining more comprehensive local and global feature information. To train and test the proposed model, a LiDAR-based worker point cloud dataset was constructed, featuring construction workers annotated with both 3D bounding boxes and 3D human keypoints. The experimental results show that the model demonstrates strong performance on the construction worker dataset with an MPJPE of approximately 0.13, providing excellent body perception capabilities essential for construction robotics.
