A Study on Part Affinity Fields Implementation for Human Pose Estimation with Deep Neural Network

This paper investigates the implementation of part affinity fields in deep neural network to estimate human body pose from images and videos. The deep neural network is capable to perform human pose estimation under various body position and activities, based on human localization and human pose detection. Human localization is inferred from the probability and affinity map calculated from the input data, while human pose detection is achieved through automated key point annotation in the affinity map cluster and skeleton generation from the detected key points. Our image-based pose estimation is conducted on several images containing single and multiple human subjects performing different activities. Our video-based pose estimation is carried on videos with different contrast conditions and different moving activities. We analyze the accuracy of automated key point annotation and its influence to the accuracy of human pose estimation.