MultiModal, Robust and Accurate Hand Tracking

3D hand tracking has a wide application prospect, for instance, in the field of human-computer interaction. Existing generative methods based on model fitting typically obtain good accuracy, but are far less robust to rapid hand motions and gestures with self-occlusions. We thus present a multimodal hand tracking system aiming to boost both accuracy and robustness. An efficient initialization strategy based on image data and inertial sensor data is designed to improve the robustness. A new optimization algorithm is proposed to improve the accuracy further by introducing gradient descent into APSO (Adaptive Particle Swarm Optimization). In addition, we construct a cylinder-cuboid hand model and establish a corresponding cost function. The proposed hand tracking method is evaluated on MSRA14 Dataset and achieves the state-of-the-art result.