<p>Human pose estimation can accurately identify and locate the key points and motion of each joint on the human body from images or videos, and abstract an arbitrary human pose skeleton, which has important application prospects in computer vision tasks. Currently, human pose estimation mainly relies on data-driven deep learning models, usually facing problems such as difficulty in training, inefficient learning, and limited generalization ability. Thereby, a physical dynamic evolution learning model for human posture estimation network is proposed to optimize the nonlinear learning property of the deep learning model using physical dynamics. The core of physical dynamic evolution learning model lies in the construction of Hamiltonian dynamical neural network evolution model (HDNE), which mainly consists of alignment network and Hamiltonian neural network (HNN). The alignment network can adapt to the output heads of any different human pose estimation models. HNN learns the Hamilton function that approximates the dynamical system to simulate and predict the evolutionary behavior of high-dimensional features for deep learning based on Hamiltonian canonical equations. Experimental results demonstrate a significant improvement in human pose detection accuracy and training efficiency under the evolutionary learning of HDNE.</p>