Research on Real-Time Motion Classification and Counting Algorithm Based on Video

High quality exercise data can often provide a basis for people to formulate scientific fitness plans. At present, the methods of motion recognition and counting through manual quantization or wearable devices with sensors are poor in convenience, visualization, efficiency and accuracy. This paper presents a real-time motion classification and counting method using computer vision technology, which makes up for the shortcomings of the previous methods. Firstly, a lightweight human posture estimation network MPE is designed to extract human bone point data, and then the data of human bone points are processed for motion classification. Finally, based on the classification results, the counting algorithm proposed in this paper is used for real-time motion counting. The counting algorithm can customize the counting standard to standardize the motion action. This method can meet the counting scene of a single person doing a variety of movements alternately. This paper mainly carries out experiments on five kinds of sports including squats, sit-ups, push-ups, pull-ups and jumping jacks. The experimental results show that the PCKh index of MPE network reaches 85.9 on MPII dataset and the AP index reaches 65.0 on COCO dataset. It also performs well in lightweight, with only 5.8G FLOPs and 5.6M model parameters. In addition, the classification accuracy of KNN classification network based on MPE is 95% on self-made dataset. The accuracy of counting is 93%, and the accuracy of counting ± 1 is 97%.