Abstract
Human pose estimation is a hot topic of computer research in recent years, which promotes the progress of society and brings many conveniences to people’s lives. Fom traditional methods to the mainstream deep learning-based methods, the primary approach in deep learning involves the use of convolutional neural networks to reduce computational complexity and improve network accuracy, but because the network structure is too deep to improve the accuracy, the trained model parameters are also very large, and it is very dependent on the input of hardware equipment. At this time, the lightweight human pose estimation can solve this problem very well. This paper mainly describes the knowledge of convolutional neural network in detail and compares it with traditional image algorithms. The OpenPose model is a classic model based on convolutional neural network that can well achieve single-person and multi-person human pose estimation model, but because the convolution kernel in its network structure is too large to increase the amount of calculation, this paper proposes three improvements to the network structure of the conventional OpenPose model. Finally, the precision of the model is improved by about 40%, which verifies the feasibility of lightweight human body posture estimation research.