Objective In recent years, real-time, accurate, and markerless recognition of sports motions has been a hot issue. The related researches on machine learning or deep learning recognition of specific sports or target motions were systemetically reviewed based on computer vision data input and a reference was provided for the application of markerless motion capture technology in the field of sports science.
Methods The documents published from January 2000 to June 2020 in 6 databases were collected, including Web of Science, PubMed, Scopus, Google Scholar, IEEE Xplore, and CNKI through Boolean logic operations. 23 research articles were included after screening, and the first author/year of publication, movement type/target action, subject information, camera parameters, image feature extraction technology, action recognition algorithm, action recognition quality evaluation method, image data training and verification methods, action recognition accuracy performance, and other key information were extracted.
Results 39% of the studies use machine learning algorithms based on support vector machines, and 35% of the studies use deep learning algorithms based on convolutional neural networks. Classification accuracy, confusion matrix, and displacement error are the action recognition quality evaluation methods used in most studies.
Conclusion Field camera settings based on computer vision, image feature extraction, and motion recognition algorithm model development need to be combined with specific sports, application scenarios, and accuracy requirements for multi-dimensional comprehensive measurement. With the development of deep learning recognition algorithms and wearable wireless sensing equipment, the accuracy, real-time, and robustness of unmarked motion capture will be further improved.
DownLoad: