Deep Learning-Assisted Fine-Grained Action Recognition From Physical Education Teaching Videos

New machine vision and computational intelligence technology in physical education (PE) are analyzed as part of an economic analysis of sports training videos. This study suggests fine-grained action recognition (FG-AR) method for examining soccer training videos. The approaches were utilized in-house for training purposes to ascertain whether a player had successfully stopped the football. The FG-AR problem is simulated using interactions between human objects (player balls). It is proposed that entity trajectories may be used as a new descriptor for detecting fine-grained sports videos. High-level semantics and human-object interactions are useful for the proposed descriptor. Second, it is recommended that the FG-AR be implemented employing a cascaded deeper communication dependent on an object-level trajectory. The computation network is constructed using a hybrid of detector and classification systems. The phenomenal effectiveness of long-term short memory (LSTM) and deep learning-based networks on sequence data processing and the high sensor effectiveness in object identification are assets to the cascaded system. The results of the experiments show that the proposed method may achieve a precision of 91%.