Camera Synchronization from Feature Points for 3D Motion Reconstruction

The thesis addresses the problem of synchronizing unsynchronized video streams from multiple cameras for 3D motion reconstruction in sports environments. Accurate temporal alignment is critical for reconstructing dynamic scenes, such as football matches, but common methods like Dynamic Time Warping (DTW) only offers frame-level accuracy, which is insufficient for high-speed motion. To solve this, the thesis implements a subframe accurate synchronization method based solely on geometric principles, without relying on machine learning or external hardware. The method estimates the temporal offset between video sequences by minimizing the Sampson error under epipolar constraints, using tracked joint positions and motion trajectories. The approach includes RANSAC for robustness and an iterative refinement scheme to improve precision. The work was carried out in collaboration with TRACAB, using real-world player tracking data. Experiments on TRACAB’s football datasets show that the method achieves temporal alignment with an average error of less than 1 millisecond and a triangulation error below 1 centimeter for offset ranges up to 0.3 seconds. These results significantly outperform frame-level methods like DTW, which introduce measurable spatial errors even under ideal conditions. The method demonstrates strong potential for real-time sports analytics and 3D reconstruction applications where hardware synchronization is unavailable. The method is highly accurate and efficient thereby, makeing it a practical and interpretable solution for vision-based multicamera systems. The runtime analysis of the proposed pipeline demonstrates strong potential for real-time execution