Human Pose Estimation Based on Lightweight Convolutional Neural Networks

Abstract

Aiming at the problems of large network model parameters and high computational complexity in human pose estimation network, an improved lightweight human pose estimation network (MobileNetV3 And Convolutional Block Attention Module Pose, MACPose) based on Lightweight OpenPose is proposed. Firstly, an improved MobileNetV3 network is introduced for feature extraction. Secondly, CBAM (Convolutional Block Attention Module) attention mechanism is integrated in the refinement stage. Therefore, human pose estimation can be carried out under the condition of reducing the model parameters and computational complexity. Under the same experimental parameters and environment configuration, the experimental results of COCO validation set show that compared with Lightweight OpenPose, MACPose reduces 16.9% in model parameters and 22.2% in computational complexity. The experimental results show that compared with Lightweight OpenPose, MACPose proposed in this paper can still achieve good results in human pose estimation with fewer model parameters and computational complexity.