View Fusion Contrast Network for Human Pose Motion Tracking

Based on the application requirements in various fields such as intelligent human–computer interaction and video surveillance, human pose and motion tracking technology has gradually become an important application in everyday life. Due to the planar nature of human pose and motion images, tracking performance is poor. Therefore, we propose a view fusion contrast network for human pose and motion tracking. First, we employ field-based methods to remove duplicate data, and use interpolation, K-Nearest Neighbors, and Expectation–Maximization methods to fill in missing data. We normalize data with mean-variance normalization to preprocess human pose and motion information and generate visual images. Then, we analyze the geometric relationships of human pose and motion images from different multi-view images using computer vision and image processing techniques, fuse multi-view images, establish a world coordinate system, convert planar information into spatial information, and extract human pose points through binary feature exclusive OR operations. Finally, we construct positive and negative samples using contrastive learning to recognize and locate similar images, thus achieving human pose and motion tracking. Experimental results show that, compared to traditional methods, this method achieves superior tracking performance in multiple evaluation metrics, including angle dispersion, intersection-over-union (IoU). The mean angle dispersion value is 0.12, the IoU is close to 0.5. These results indicate that this method holds practical application value for human pose and motion tracking.