Machine Vision-Based Smart Basketball System and Hardware Implementation

To address the challenge of acquiring basketball shooting parameters in a non-contact and quantitative manner, this paper proposes a machine-vision-based prototype system for intelligent basketball shooting feedback. The system constructs a basketball-and-hoop detection dataset for YOLOv5-based object detection, enabling the recognition of basketball and hoop targets in shooting scenarios. By integrating a Semi-Global Block Matching (SGBM) stereo-ranging algorithm, the system obtains the three-dimensional (3D) coordinates of the basketball and the hoop. A shooting evaluation and feedback module is then developed based on an ideal projectile-motion model to analyze the release position, release angle, and release velocity, and to provide reference release velocity, reference force, and make/miss feedback under the ideal model. To improve portability, the complete workflow, including object detection, stereo ranging, and feedback generation, is deployed on an NVIDIA Jetson Nano B01 development board, with TensorRT employed for inference acceleration. Experimental results show that the trained detection model achieves a frame-level recognition rate of about 95% for basketball and hoop detection in the test video after dataset expansion, and that the stereo-ranging module provides effective depth estimation within the scale of the shooting experiments. The system enables trajectory extraction and release-parameter estimation for typical shooting clips. The average relative difference between the velocities extracted by the system and those calculated using the ideal projectile-motion model is 1.96%. In additional experiments involving multiple players, shooting distances, and illumination conditions, the system also shows a certain degree of robustness. Jetson Nano tests further show that the complete TensorRT YOLOv5s and SGBM workflow achieves an average processing rate of 3.65 FPS, an average per-frame latency of 261.31 ms, and an active-stage board power consumption of 5.41 W. Overall, the proposed prototype can provide non-contact, reference-oriented quantitative feedback for basketball shooting training, offering a feasible solution for applying machine vision technology to basketball training assistance.