3D human pose estimation based on multi-feature extraction

As an important task in the field of artificial intelligence and computer vision, 3D human pose estimation has received widespread attention and has produced many applications in fields such as human-computer interaction and movie game production. However, 3D human pose estimation still faces significant challenges, mainly including human occlusion issues and dataset perspective redundancy issues, which seriously affect the accuracy and speed of 3D human pose estimation results. This paper proposes a multi-feature extraction method for 3D human posture estimation. Firstly, by collecting image data from multiple camera perspectives, the collected image data is placed into a 2D human pose estimation network model to obtain 2D joints. Then, the collected human data is input into a joint confidence calculation network model to obtain the weight values of each joint point in the perspective image, Subsequently, the 2D human joint heatmap is calculated using a heatmap weight calculation network to calculate the weight of the heat map, and the weighted 2D human joint heatmap is obtained by fusing the weight features under various views. Finally, the weighted 2D human joints heatmap and the weight values of each joint point in the perspective image are input into the triangulation algorithm to map the 3D human joint points in space. The key idea of this paper is to design a joint confidence calculation network to learn the confidence weights of each joint from the input image, and extract a confidence matrix that reflects the quality of the heatmap to improve the quality of the heat map features in the occluded view. In addition, a perceptual hash algorithm is used to perform a de-view experiment on the Occlusion-Person dataset, which improves the model inference speed while ensuring the accuracy of the results.The method in this paper is end-to-end differentiable, which can significantly improve the efficiency and robustness of the algorithm. This paper evaluates the method using the Mean Per Joint Position Error(MPJPE) metric on two common datasets, Human3.6M and Occlusion Person, achieving results of 27.3 mm and 9.7 mm, respectively. Experimental results show that the performance of this algorithm has been significantly improved compared to the most advanced methods.