Animation Pose Generation Model Based on Kinect Depth Image and Occlusion-Robust Pose-Maps Algorithm

Aiming at the problem of low efficiency and difficulty in adapting to complex interactive scenes in traditional animation pose generation methods, a dynamic pose generation model based on Kinect depth image and occlusion robust pose map algorithm is proposed. Constructing a 3D skeleton sequence through spatiotemporal sparse sampling, effectively repairing missing areas in depth data using matrix filling techniques, and ensuring motion smoothness through dynamic median filtering. The model innovatively integrates Kinect depth information with monocular visual pose estimation results, utilizing an improved joint point representation method and Kalman filter data fusion strategy to significantly improve robustness in occluded scenes while maintaining spatiotemporal consistency. The experiment showed that the improved model achieved correct limb percentage values of 87.7% and 86.8% on the Human3.6M and MPI-INF-3DHP datasets, respectively. In addition, the improved model performed better than the comparison model on both datasets in terms of the percentage of correct keypoints and the percentage of detected joints, reaching (75.4%, 74.6%) and (83.4%, 82.3%). In the evaluation of registration error, the research model showed the lowest error on both datasets, and the error values did not exceed 2. In addition, in the case analysis, the angle fluctuation curve presented by the research model was most in line with the actual situation. In the visual evaluation of motion capture effects, the animation generated by the model had smooth poses, reasonable hand and leg movements, ergonomic joint movements, and no abrupt jumps or disjointed frames, resulting in a better overall effect. Overall, the proposed model can be effectively applied in animation production, providing a new technological means to enhance the quality of animation production.