Premium
Sport action recognition by fusing multi‐source sensor information
Author(s) -
Shi Jizu
Publication year - 2021
Publication title -
internet technology letters
Language(s) - English
Resource type - Journals
ISSN - 2476-1508
DOI - 10.1002/itl2.279
Subject(s) - computer science , classifier (uml) , artificial intelligence , kalman filter , action (physics) , computer vision , pattern recognition (psychology) , quantum mechanics , physics
In traditional sport and training, the quality of training plan mainly relies on individual observation and experience of coaches, which inevitably depends on individual objective opinion. With the emerging of sensors and body area network (BAN), it becomes possible to automatically recognize athletes' posture. Thus, the coaches can make decision based on recognition results to enhance athletes' competitive ability, which can greatly improve the effectiveness and quality of the sport training. This paper proposes an automatically action recognition system for sport training. First, we collect body information from body sensors. Second, the information from multi‐source sensors is input into a Kalman filter to remove the interference noise. Third, the denoised signals are divided as several unit actions according to four limbs. Fourth, the unit actions are represented as a matrix consisting of acceleration, angular velocity, combined acceleration, and combined angular velocity. We extract frequency domain features and time domain features to represent the matrix. Lastly, the extracted features are used to train an intelligent classifier which is used to predict the future postures. The experimental results on a sport training dataset demonstrates the effectiveness of our framework for action recognition for sport training.