Sanshou is becoming more and more popular as a comprehensive sport that integrates traditional Chinese martial arts and modern fighting techniques. In this paper, a CNN-LSTM model fused by a convolutional neural network (CNN) and a long-short-term memory network (LSTM) is proposed, which is capable of capturing global information and sequence relationships to realize accurate and real-time recognition of sparring techniques. A dynamic time planning algorithm is utilized to calculate the difference in movement joint angles between the standard and test movement sequences, and a movement evaluation formula is defined to evaluate the sparring movements, which provides intuitive and reliable training suggestions for the trainees, thus improving their learning efficiency. The research results show that the CNN-LSTM model can recognize six kinds of sparring actions with 98.69% recognition accuracy, while the action scoring model proposed in this paper can realize the effective evaluation of sparring actions and achieve fair and impartial auxiliary scoring. Finally, the sparring tactics confrontation decision is proposed to improve the tactical quality and effect of the students during the competition.
