Human pose estimation based on intra-level feature fusion

Abstract

To improve the low estimation accuracy and poor robustness of invisible human joints in 2D human pose estimation in computer vision, a convolutional network based on intra-layer feature fusion is proposed to improve the estimation accuracy. Firstly, the convolutional block based on intra-layer feature fusion is proposed to obtain abundant local features and reduce the influence of local information loss. Secondly, intermediate supervision is introduced to reduce the influence of vanishing gradient in deep neural network. Finally, several proposed convolutional blocks form the neural network in cascade, through continuous Up-Sampling and Down-Sampling operations, convolutional blocks obtain local information from images with different resolutions and have the information fully fused, the local information after fusing can establish the connection between joints and accurately estimate the joints. The proposed neural network achieves first-class results on standard benchmark including the LSP dataset and its extended dataset.