Human Action Recognition and Counting Based on Embedded System

Although action recognition and counting based on embedded systems have already emerged in daily life, such as using mobile phones for exercise training, it is still a challenging problem to balance the fluency of operation in hardware systems and the accuracy of counting. A hardware platform with built-in acceleration modules and a moderately sized deep neural network may be a feasible solution. In this paper, the NVIDIA Jetson TX2 and the CNN model are selected to build a realtime application. Then, the feature extraction network and branch network are compressed using a lightweight method for better adaptation. Thus, whether it is a still posture or active movement, the key coordinate of the human body joint can be obtained, and then the actions can be distinguished using corresponding discrimination methods. Especially in the case of sports, such as Push-ups and Sit-ups, the angles of joints and bodies can be calculated mathematically and the number of actions can be estimated according to the action standards. The experiments show that the proposed algorithm can achieve 13~16 frames per second(FPS) and the accuracy of movement recognition and counting is higher than 90% in the choosen embedded system.