Model Optimization Methods for Real-time 2D Human Pose Estimation Applied to Robots: A Survey

With the development of technology, there are more and more robots around us. These robots need to recognize human pose to monitor people or interact with people, so the human pose estimation model is required to be deployed on robots to achieve these functions. Human pose estimation involves the processing of more than a dozen joints, so the human pose estimation model usually has a relatively large amount of calculation resulting in a relatively slow detection speed. Transferring these large amounts of image data to cloud servers for human pose estimation detection will consume a lot of network resources and detection time will be delayed. And the robot’s local computing power is limited. So it is difficult to deploy both human pose estimation algorithm models and other complex algorithm models locally on robots. This requires various algorithms running locally on the robot to consume as little computing resources and real-time as possible. This paper studies the optimization methods of real-time 2D human pose estimation models studies in recent years, which are divided into three aspects: the optimization of the feature extraction network, the optimization of the joint point detection and the optimization of the joint point connection. And summarized the advantages and disadvantages of these optimization methods, attempting to apply them to robots.