Improved YOLOv11 for Multi-class Vehicle Detection from the Perspective of Unmanned Aerial Vehicles

In intelligent transportation systems, unmanned aerial vehicles (UAVs) provide a unique aerial perspective for traffic monitoring and incident response, offering wide-area coverage that is difficult to achieve using traditional methods. However, vehicle detection in UAV imagery faces challenges including significant size variations, limited small-vehicle details, complex background clutter, and constrained onboard computing resources, necessitating efficient yet accurate detection models. To resolve these challenges, the present research has designed an enhanced YOLOv11 network optimized for UAV scenarios. The BiFPN mitigates scale-variant degradation through adaptive multi-scale fusion, addressing persistent blind spots in micro-object detection. A high-resolution P2 feature layer with deformable convolution further enhances small-object spatial localization in the detection head. Additionally, the Focal-ECIoU loss function addresses scarce high-quality regression samples through dynamic gradient modulation, prioritizing promising candidate boxes to mitigate localization bias and accelerate convergence. Validation on a curated UAV dataset yielded AP@50 and mAP values of 45.6% and 32.0%, respectively, representing a 7.5%/7.6% advancement over YOLOv11 baseline, while maintaining approximately 8.4 million parameters. These results demonstrate the significant advantages of balancing accuracy and efficiency for aerial vehicle detection.