Research on Unmarked Motion Action Recognition Technology Based on Computer Vision
Lun Zhao · 2024
The main purpose of this study is to explore the issues of real-time, accurate, and unmarked recognition of sports movements in recent years. By reviewing the relevant research on machine learning or deep learning for specific sports or target actions based on computer vision image data input, the aim is to provide references for the application of unmarked motion capture technology in the field of sports motion recognition. The research employed a literature review methodology, conducting searches in six databases, namely Web of Science, PubMed, Scopus, Google Scholar, IEEE Xplore, and China National Knowledge Infrastructure (CNKI), covering publications from January 2000 to June 2020. Through boolean logic operations on the retrieved literature, key information such as first author/publication year, types/targets of motion, participant information, camera parameters, image feature extraction techniques, action recognition algorithms, evaluation methods for action recognition quality, training and validation methods for image data, and performance metrics for action recognition were extracted. After screening, a total of 23 articles were included in the study. The findings revealed that $39 \%$ of the studies utilized machine learning algorithms based on support vector machines, while $35 \%$ employed deep learning algorithms based on convolutional neural networks. Commonly used evaluation metrics for action recognition quality included classification accuracy, confusion matrix, and displacement error. The development of computer vision motion capture, models, and algorithms demonstrated promising applications in areas such as action technique recognition and sports performance analysis. Traditional machine learning algorithms like support vector machines and principal component analysis remain dominant in action recognition technology; however, in certain scenarios, the performance of deep learning algorithms surpassed that of traditional machine learning methods. The development of computer vision-based field camera setups, image feature extraction, and action recognition algorithm model requires a multidimensional assessment considering specific sports, application scenarios, and accuracy requirements. With the advancement of deep learning recognition algorithms and wearable wireless sensing technology, the precision, real-time capability, and robustness of unmarked motion capture are expected to further improve.