Search

In understanding human behaviour, computers sometimes condense their analysis of human behaviour into an analysis of the state of movement of keypoints in the human body. Thus, the technique of human pose estimation provides a convenient means for machines to recognise people's behaviour. The lightweight pose estimation network enables computers to detect human poses in real-time. This paper proposes Lightweight HRNet, a bottom-up lightweight network for multi-person human pose estimation. The network is proposed based on the HRNet architecture and includes four network branches with different resolutions and two network stages. The network backbone uses the ShuffleNet model, which allows the network to be better used on smaller devices. Notably, Lightweight HRNet focuses on a problem with multi-resolution, multi-branch parallel networks: not all stages of the network contain information about the feature maps of all its branches. Therefore, we proposed the Channel Exchange Module (CEM), which improves the exchange of information between each stage of the network and each of its branches. Among the tasks of human pose estimation, the addition of CEM improved the network accuracy of Lightweight HRNet by 0.9% in the COCO2017 test-dev. Finally, the network was able to achieve 46.6% accuracy on the COCO2017 test-dev. This accuracy is superior in the study of bottom-up lightweight human pose estimation networks. [ABSTRACT FROM AUTHOR]

Aiming at the problem that human behavior is difficult to identify, a bottom-up and horizontal connections Cross-dimensional Feature Fusion MLP model (CFF-MLP) was proposed in this paper. To improve the detection performance of human behavior and realize the effective capture of continuous behavior features, a human key point Vector tracking (VT) algorithm based on optical flow was designed in the CFF-MLP model. This model can obtain the behavioral features and local dependencies of different layers hierarchically. Furthermore, it enhances the generalization ability of the model. The validity of the proposed model is verified and analyzed using KTH dataset and compared with other research methods. The experimental results show that the CFF-MLP model can effectively improve the detection ability of human micro-motion, and its average accuracy rate reaches 94.23%. It improves the detection efficiency of human behavior. [ABSTRACT FROM AUTHOR]