NOC-YOLO: An exploration to enhance small-target vehicle detection accuracy in aerial infrared images

Infrared imaging technology enhances the visibility of vehicles in the dark from an aerial perspective. However, the loss of texture and boundary features of targets in infrared images impedes the accurate detection of small-target vehicles. To address this issue, we propose Nocturne-You Only Look Once (NOC-YOLO), using YOLOv10n as the baseline model, aimed at improving the detection accuracy of small-target vehicles in drone-captured infrared images. We introduce three key improvements to the baseline model, incorporating attention mechanisms and multi-scale feature fusion to enhance the interaction between features. Extensive experiments on the DroneVehicle dataset demonstrate NOC-YOLO’s superiority: it achieves state-of-the-art mAP50 (79.5 %) and mAP50-95 (63.2 %), surpassing YOLOv10n by 1.4 % in mAP50, respectively, with only a 10 % increase in parameters and 12 % higher GFLOPs. Compared to non-YOLO methods, NOC-YOLO achieves significant performance improvements while maintaining exceptional lightweight characteristics, making it highly suitable for real-time deployment in resource-constrained environments. This balance of high precision and efficiency provides robust technical support for safety–critical applications like nighttime driving trajectory extraction and collision risk analysis.