BSTR: a transformer-based framework for robust and efficient basketball object detection
Lezhong Sun, Xiaopeng Lin, Zhi Li · 2025
With the continuous advancements in computer vision and machine learning technologies, automated systems have achieved significant progress in the field of basketball sports analysis. However, traditional object detection methods often struggle to extract stable features under complex shooting conditions such as illumination variations, motion blur, and occlusions. Furthermore, the intricate ball trajectories and dynamic athlete postures in basketball games further increase the difficulty of object detection. To address these challenges, we propose an end-to-end basketball object detection framework named BSTR (Basketball Sports Transformer), designed to achieve efficient and accurate object detection in complex scenarios. The BSTR framework consists of a Consistency Augmentation Module (CAM) and a Motion-Perceptive Contrast Enhancement Module (MPCE). CAM guides the model to learn in multi-scale feature spaces and ensures the extracted features exhibit enhanced stability and robustness by constraining prediction consistency under various environmental perturbations. MPCE integrates spatiotemporal features with dynamic attention mechanisms to precisely capture and align critical motion information across adjacent frames, thereby improving the model's target recognition capability in complex motion patterns. Experimental results demonstrate that the BSTR framework outperforms existing methods on multiple basketball game video datasets and exhibits superior robustness.