The research and application of motion capture technology have always attracted much attention. This paper sorts out the current research status and analyzes the limitations of the existing technologies. The algorithm models were mainly studied, including two-dimensional pose estimation, three-dimensional pose estimation and spatio-temporal optimization networks, etc. By combining the physical constraints of human bones and anatomical prior knowledge, the network was guided to learn more reasonable three-dimensional poses and reduce unnatural pose estimation results. Extensive experimental verification was carried out on public datasets. The results show that the proposed algorithm is superior to the existing mainstream methods in aspects such as two-dimensional pose estimation, three-dimensional pose estimation and time series optimization. Especially in complex environments and challenging scenarios, it performs outstandingly, improving the accuracy of human key point detection, solving problems such as expensive equipment, high environmental requirements and difficulty in handling occlusion existing in traditional methods, lowering the threshold of motion capture, and broadening its application scenarios