Badminton actions detection from sensor data based on Deep Belief network optimized by Advanced Snake optimizer

Abstract

Badminton is a kind of racquet sport which has fast and accurate movements that highlights the need for badminton action’s automatic identification within training, sport analysis, and coaching. However, the identifying actions of badminton from sensor data is difficult according to the complex nature of human actions. Such movement identification systems focus on general activities like sitting, walking, and running rather than actions that are badminton-specific. Also, the badminton players’ sensor data show different measures that makes the traditional feature normalization methods useless. This study presents a new approach for badminton action identification through use of DBNs or Deep Belief Networks, which are optimized by the ASO or snake optimizer’s improved version. The proposed DBN/ASO model is tested on the Badminton Sensor Dataset (BSS), which includes 25 players performing 10 types of strokes (1,140 samples). The model performed satisfactorily in experiments, achieving 93.2 % accuracy, 94.1 % sensitivity, 92.3 % precision, 93.8 % specificity, 93.2 % F1-score, and a Matthews Correlation Coefficient (MCC) of 0.915, surpassing CNN/LSTM, MM-AGNES, NDT-GCN, 3D:VIBE, and Multi-Sensor (M−S) state-of-the-art model performances. AUC = 0.92 from the Receiver Operating Characteristic (ROC) analysis, confirms its strong discriminative ability. Comparatively benchmarking on CEC test functions ASO has yielded the mean best fitness 9.12e⁻⁷ on F1 (Sphere) and 1.43e⁻³ on F2 (Rosenbrock), associated with the lowest standard deviation across all functions, thus demonstrating better convergence and robustness. This substantiates the efficacy of the proposed framework for the accurate recognition of complex badminton actions from wearable sensor data, in turn laying down a pathway for intelligent coaching and real-time performance analytics.