Traffic Multi-Object Detection Method Based on YOLOv10n

To address the challenges of balancing detection accuracy and real-time performance in complex traffic scenarios for vehicle-mounted embedded platforms and road monitoring, this paper proposes YOLOv10n-FTAS, an optimized lightweight detection framework based on YOLOv10n. The main innovations include: (1) Designing a C2f-Faster-EAMA module in the backbone network that enhances feature representation through channel-spatial cooperative attention mechanisms; (2) Proposing a novel statistics-enhanced attention mechanism (Token Statistics-enhanced PSA, TS-PSA) by integrating Token Statistics Self-Attention; (3) Constructing a Dynamic Sample-Attention Scale Fusion module (DS-ASF) that achieves multi-scale feature fusion through deformable convolution and adaptive sampling strategies; (4) Adopting Shape-IoU loss function with geometric constraints to optimize bounding box regression. Experimental results demonstrate: The improved model reduces parameters and computations to 5.5M and 5.8G respectively, representing 5.17% and 13.4% reductions compared to the baseline. It achieves 90.3% precision, 92.5% mAP@50, and 70.6% mAP@50:95%, showing improvements of 2.15%, 4.52%, and 2.63% respectively. This solution effectively resolves detection deviations in dynamic complex scenarios while providing high real-time performance.