Construction of 3D Human Motion Capture Model Driven by Multimodal Sensor Fusion

The construction of 3D human motion capture models often relies on a single sensor, resulting in incomplete data and low motion capture accuracy. Therefore, this paper studies a multimodal sensor fusion-driven method for constructing 3D human motion capture models. Data from different types of sensors is integrated, acquisition points are rationally arranged, and Kalman filtering is used to fuse the data to accurately estimate the human motion state. Human posture matching is performed by segmenting the fused data sequence into frames to enhance the capture capability of key spatiotemporal features. The DTW algorithm is then used to measure the similarity of posture sequences. Using the fused data and posture information, a 3D human motion capture model is constructed to capture joint movement changes. Experimental results show that under normal and low light conditions, this method correctly captures more samples than comparison methods for seven types of movements. In joint trajectory comparison, the simulated hip, knee, and ankle joint angles are close to the actual values, with an average absolute error of less than 1°. On the IMHD2 multimodal dataset, the accuracy of this method exceeds 97% for movements such as punching, demonstrating superior performance in motion capture accuracy, joint trajectory simulation, and movement classification.