Multi-View Camera-Based Large-Scale Multi-Person Motion Capture and Recognition

This study aims to address the challenge of achieving high-precision 3D multi-person motion capture and recognition in large-scale, complex, and highly interactive environments. We propose and implement a real-time motion capture system based on multi-view cameras that provides high-quality 3D skeletal data to support downstream human action recognition tasks. The system adopts a coarse-to-fine two-stage architecture. First, 2D keypoint heatmaps from multiple camera views are back-projected into a shared 3D voxel space to construct a unified volumetric feature representation. Then, through orthogonal projection, a lightweight 2D convolutional network accurately localizes human centers and regresses 3D bounding boxes. Based on these boxes, the network extracts local volumetric features, fuses information from three orthogonal planes, and regresses precise 3D keypoint coordinates. To ensure temporal identity consistency, the system integrates a multi-target tracking strategy combining the Hungarian algorithm and appearance-based Re-ID features. Additionally, keypoint interpolation and temporal smoothing techniques are employed to enhance the continuity and stability of motion sequences. Experiments conducted on our self-built 100 [Formula: see text] test platform equipped with 10 synchronized cameras demonstrate that the proposed system achieves 95.21% AP, 80.19% 3D PCK, and 46.31[Formula: see text]mm MPJPE on our challenging internal dataset — surpassing state-of-the-art methods such as VoxelPose. Furthermore, the high-precision skeletal data captured by our system show strong potential in facilitating human action recognition under complex multi-person interaction scenarios, providing a robust foundation for intelligent motion analysis, sports training, and human –computer interaction applications.