Abstract:With the development of mobile and embedded devices, higher demands are placed on 2D human skeletal keypoint detection networks. Designing lightweight neural networks is an important approach to solve the problem of large number of network parameters and large computational effort. First, the mainstream methods and lightweight neural networks for 2D human skeletal keypoint detection based on neural networks are introduced; then, the lightweight human pose estimation methods based on neural networks in recent years are classified and summarized, and the 2D skeletal keypoint detection methods are grouped into four categories according to the lightweight way of neural networks: lightweight backbone networks, deep separable convolution, Dense connection mechanism and Lightweight Bottleneck, and analyzed their advantages and disadvantages and lightweighting means; finally, the common data sets and corresponding evaluation metrics are introduced, and the improved lightweighting methods are compared with experimental data. A summary and outlook are given in relation to the current challenges and future development trends of the research.