Top-Down Meets Bottom-Up for Multi-Person Pose Estimation

Abstract

Body posture, as a basic feature of humans, occupies an important position in human-computer interaction and other fields. In order to accurately recognize human body posture features and eliminate misrecognition caused by different human movements. This paper proposes a human body posture detection method that combines top-down and bottom-up two-way. This method divides the input image into two forks, and uses the Mask R-CNN model to extract key point features and human target location features. The key point recognition feature map is divided into two, and using CNN to extract Part Affinity Fields and Part Confidence Maps. The obtained confidence map is added with the human body position feature map. Then use PAFs to segment the key points of different individuals in the result. Finally, the Hungarian Algorithm is used to process the maximum matching of the bipartite graph to obtain the human body posture detection result. Experimental results show that this method can accurately recognize human posture. It has good performance on different data sets and eliminates the estimation accuracy defects of traditional models due to the complex high-dimensional motion of the target.