Distributed Human Pose Estimation Based on BlazePose

Abstract The lightweight of human pose estimation model is one current research hotspot in the field computer vision. emphasis to promise while improving network performance. In view computing and processing performance monocular static image method deployed mobile devices, a distributed was proposed. framework constructed by simplifying BlazePose as core, 5×5 convolution kernels were run for feature downsampling multi-thread simulated environment. power output devices reduced simulation. Experimental results showed that proposed had 3%~11% improvement compared with other models simulation Compared traditional model, reconstructs neural deployment framework, thus indirectly enabling stable recognition equipment under low power.