Deep Learning-Based Involution Feature Extraction for Human Posture Recognition in Martial Arts

With the development of computers in recent years, human body recognition technology has been vigorously developed and is widely used in motion analysis, video surveillance and other fields. This study is based on deep learning to improve human pose estimation. Firstly, Involution's feature extraction network was proposed for lightweight human pose estimation, and this feature extraction network was combined with existing human pose estimation models to recognize human pose. Label and classify each joint point of the human body separately, add weights to each different part, extract feature between joint points at different times, and then input the extracted feature into long short-term memory neural networks for recognition. The experimental results show that the improved human pose estimation model reduces the parameter and computational complexity by about 40% compared to the original model, while also slightly improving accuracy. Comparing the performance of models under various algorithms with the proposed model in this study, the accuracy under the Eigen method is 81.3%, the accuracy under the STOP method is 82.5%, the accuracy under the DMM&HOG method is 85.3%, the accuracy under the Actionlet method is 87.6%, and the accuracy under the JAS&HOG2 method is 83.5%. The accuracy of the InNet LSTM method is 90.6%. The results indicate that the proposed model has good performance and can recognize different martial arts movements.