In view of the current situation that sedentary people lack of exercise for a long time, resulting in sub-health state and the lack of supervision of existing training methods, this paper proposes a method of human motion recognition and counting to realize the accurate recognition and counting of four kinds of training without equipment. Taking the video information of the trainer captured by the mobile camera as the input, the human skeleton point data processed by the BlazePose network model is processed by data filtering and feature extraction, and three common machine learning algorithms are used for action classification. The classification results are combined with the bone information, and the peak and trough counting algorithm is used to count the number of training actions completed. The experimental results show that: using GBDT classification algorithm, the action recognition rate is 96.5%, and the accuracy of counting algorithm is 98.9%, which has good practical application value.