This paper recognizes the human motion in martial arts movements based on the LSTM recurrent neural network algorithm, which uses signal statistics to extract the action time series features through the data collection from the relevant human motion sensors in order to effectively recognize the athletes’ movements. The parameter settings of the LSTM neural network model are optimized using an improved particle swarm algorithm. The experimental process is then created. After filming the video, it is divided into two documents, A and B, and saved. The experimenter’s actions undergo technical analysis using LSTM. Three groups of athletes were invited to carry out the actual test after the design was completed, and biomechanical analysis was performed based on the experimental data. The experiments showed that the activation of the left medial femoral muscle of the athletes was greater than that of the left lateral femoral muscle, and the root mean square of electromyography was 25.38±11 and 27.54+11, respectively. The angle of the left leg was 69±2.5, and that of the right leg was 71±9.2, which was more than 80% of the safe rotational amplitude. It is easy to cause sports injury if there is too much difference in the activation level of leg muscles during whiplash.