Enhancement of accuracy and analytical efficiency of gymnastics video using convolutional neural network and semantic analysis approach

Gymnastics competition's conventional video analysis approach mostly depends on shallow image processing technology, which is challenging to sufficiently capture the subtleties of participants' motions and semantic information in the competition. This work suggests a semantic analysis approach of gymnastics competition film based on domain knowledge and depth features to compensate this deficit. First, one builds a knowledge base based on action classification in addition to the domain knowledge of gymnastics competition. Second, the temporal and spatial traits of video and the dynamic performance of athletes are extracted using convolutional neural network (CNN) in combination with long-term and short-term memory network (LSTM). In the analysis of gymnastics competition video, the experimental findings reveal that this approach produces better recognition accuracy and analytical efficiency than the conventional one.