Pose-Aware Convolutional and Recurrent Framework for Real-Time Basketball Shooting Assistance and Training

From the past few years, traditional basketball training depends heavily on subjective coach observation which limits objective skill assessment and consistent feedback. Even though, previous researchers have explored the existing deep learning models for action recognition, still there are challenges with real-time adaptability, illumination variation and lack of biomechanical awareness. Therefore. This manuscript introduces a lightweight and real-time basketball shooting assistance system which is a Pose-Aware Spatiotemporal Network for real-time basketball shooting assistance named as PAST-Net. In particular, to overcome the challenges, the proposed approach integrates a pose-aware hybrid Convolutional Neural Network (CNN)-Gated Recurrent Unit (GRU) network with multimodal feature fusion. Subsequently, the video frames from the BASKETBALL-51 dataset are processed which helps to extract optical flow for motion representation and skeletal key points for body posture estimation. Further, the CNN module captures spatial features, while the GRU models temporal dynamics of the shooting sequence. Finally, the fused features are categorized into Correct, Incorrect, or Needs Correction categories which allows automated evaluation of player performance. Additionally, a novel pose-aware feedback mechanism is incorporated which provides real-time corrective cues based on deviations from an ideal shooting template that assists the players in refining elbow positioning, wrist release and stance alignment. Thus, the proposed PAST-Net obtained accuracy of 94.5%, precision of 94.3%, recall of 94.8% and 94.7% F1-score, when compared with existing model.