DeepGaitLab: Accurate and Flexible Markerless Motion Tracking Powered by Synthetic Data

Scalable and accurate human motion tracking is expected to modernize the diagnosis and prognosis of gait pathologies, sports performance optimization, and human movement research at large. While recent advances in computer vision are promising, innovation has primarily focused on single-view approaches, which are convenient and could be applied to videos from commodity devices, such as smartphones. However, a range of biomedical applications require higher accuracies. Current multi-view tools either lack sufficient accuracy to justify the added burden of camera calibration or require higher-density multi-camera setups that discourage adoption in out-of-laboratory settings. Here, we present DeepGaitLab, an open-source framework that yields accurate three-dimensional (3D) kinematics while operating flexibly across a range of camera configurations, including only two, and foregoes the time-consuming step of inter-camera calibration. Trained on large synthetic data, DeepGaitLab overcomes the accuracy and generalizability constraints of current tools relying on real data. It outperforms both commercial and open-source alternatives and exhibits monotonic accuracy improvement with additional cameras. Unlike existing systems, its accuracy does not degrade when applied to individuals with mobility limitations. We evaluated DeepGaitLab in 80 individuals, including healthy adults, individuals recovering from anterior cruciate ligament reconstruction, individuals recovering from stroke, and ones with mild cognitive impairment, captured in two distinct environments. In addition to outperforming existing tools and demonstrating utility across three clinically distinct populations, DeepGaitLab offers the optional feature to enhance accuracy via an environment-specific fine-tuning strategy without requiring new labeled data. Together, these advancements establish DeepGaitLab as a practical and scalable platform for real-world deployment, bridging the gap between research-grade biomechanics, emerging artificial intelligence (AI) tools, and clinical impact. All the data, code, and trained models are publicly shared.