Multi-Person Pose Estimation Based on MobileNet Neural Network
Xianfen Xie, Tianrui Huang · 2023
Human pose estimation has always been a hot issue in the field of computer vision. Its main content is to let the computer locate the key points of the character (also called joint points, such as elbows, wrists, etc.) from images or videos. Inferring the poses of multi-person in an image is a unique set of challenges. Human pose estimation is mainly used in fields such as motion recognition, human-computer interaction, intelligent security, and augmented reality. The main problems of human pose estimation include: (1) The image may contain an unknown number of people, who can appear in any position or on different scale; (2) Due to contact or body weight, the interaction between people will cause complex spatial interference, making the association between components difficult; (3) At runtime, complexity tends to increase as the number of people in the picture increases. This problem makes real-time performance a challenge. The main function of this paper is as follows: aiming at the shortcomings of the previous human pose estimation algorithms in model size and computing speed, this paper proposes a new network structure and a bottom-up method, using MobileNet portable network for image feature extraction, using joint points confidence field and part affinity field are used to predict the position of the joint points of the human body and connect the joint points, and use the sparse matrix storage method to filter and store the vectors in the calculation process to complete a base model for bottom-up human pose estimation with better computing speed, referred to as MBHPE.