Abstract
Intelligent training systems are becoming essential today in sports science in the context of improving athletic performance and injury prevention. Nevertheless, most action-recognition systems are based on traditional deep learning models in terms of their representational ability to distinguish biomechanically similar football actions, and the hyperparameter optimization is frequently achieved through manual trial and error or simple metaheuristics (e.g., PSO, GA), which is slow to converge, inefficiency in high-dimensional space and poor exploration-exploitation balance. To seal these loopholes, this research work suggests a new intelligent strength training model of a football player that combines an improved motion discrimination ResNeXt convolutional neural network, provided by cardinality-based feature learning with an Upgraded Chimp Optimization Algorithm (UCOA), which improves global search efficiency with chaotic map start and elimination step to prevent local optima. The system is tested on the Berkeley MHAD dataset, where it automatically recognizes football-relevant motions (e.g. kicking, jumping, squats) and visualizes them to specific muscle groups so that they can be recommended specific strength exercises. Experimental findings indicate that the UCOA-optimized ResNeXt attains a classification accuracy of 93.7% and an F1-score of 92 and is more accurate than the traditional deep learning models and hybrid optimization baselines.