Real-time Detection of Basketball Free-throw Technique Movements Using a Deep Learning Model Based on Smartphone Video

Abstract

This study aims to develop a deep learning model based on smartphone video recordings to enable real-time detection and feedback of basketball free-throw technique execution. Traditional analysis methods rely on manual observation, which suffers from feedback latency and high subjectivity. In this research, free-throw videos of 10 athletes were recorded using smartphones, and certified coaches classified the videos according to standard technical criteria; error categories included unstable stance, improper knee flexion, and excessive elbow abduction. The MoveNet model was employed to extract 17 key body points, and 26 features—including joint angles and inter-joint distances—were computed. The datasets were enhanced using techniques such as Gaussian noise, jitter, scaling, translation, and temporal distortion. Four models were trained and compared; on the balanced dataset, the dual-stream convolutional neural network model achieved the best overall performance, with an accuracy of 93.8%, precision, recall, and F1-score of 93.2%, 92.5%, and 92.2%, respectively, and a logarithmic loss of 0.23. Five-fold cross-validation yielded an accuracy of 93.8% ± 0.5%, demonstrating excellent stability. Confusion matrix analysis revealed that the model correctly predicted 804 out of 876 samples, exhibiting optimal class balance. This study validates the feasibility of this approach in providing athletes with instantaneous quantitative feedback; future work may explore the integration of edge computing and multimodal fusion to enhance robustness.