ウェアラブルスポーツ画像における高速小ターゲット検出のための動的スパースアテンションを備えたYOLC

To address the challenge of balancing high-speed, small-target detection accuracy and real-time performance in wearable devices under stringent resource constraints, this paper proposes a lightweight integrated detection and tracking framework, YOLC (You Only Look Clusters). This framework is based on MobileNetV3 as the backbone, integrating depthwise separable convolution and cross-stage partial connection structures to build a lightweight feature extraction network. A dynamic sparse spatial attention mechanism is designed, which adaptively generates spatial weights through local texture variance response and activates only the top 30% of high-response regions, thereby achieving efficient scheduling of computational resources and suppressing background interference. The framework further proposes an attention-guided bidirectional weighted feature pyramid to enhance the feature fusion effect for multi-scale small targets. A coordinate attention module is integrated into the detection head to improve positioning accuracy, while lightweight optical flow estimation and a confidence-adaptive matching strategy are combined to optimize the ByteTrack association process and improve trajectory stability under high-speed motion. Experimental results demonstrate that the proposed method achieves a mAP@0.5 of 75.3% and a small target recall rate of 82.4 ± 1.1%, with an inference processing speed of 53.5 FPS. It significantly improves the robustness of high-speed, small target detection and tracking in typical motion scenarios.