IMV-YOLO: infrared multi-angle vehicle real-time detection network based YOLOv11 for adverse weather conditions

Purpose In adverse weather, infrared imaging is affected by wavelength and atmospheric attenuation, resulting in images with missing textures, low resolution and insufficient contrast. These defects significantly increase the probability of missed detections and false detections in vehicle detection. Furthermore, existing methods are only applicable to fixed detection perspectives and struggle to handle the variations in vehicle thermal distribution across different viewing angles. Design/methodology/approach This research first established a dataset named AIR-MAV (adverse weather infrared multi-angle vehicles) for multi-angle infrared vehicle data in adverse weather conditions. Secondly, the C2-AIFI model was innovatively proposed, effectively improving the accuracy and noise robustness of infrared vehicle detection. In order to improve the processing capability of the network for different-scale features, channel shuffle and one-shot aggregation strategies are adopted to strengthen cross-channel feature interaction. Finally, we introduced the DHead-Detect module, which integrates three attention mechanisms to focus on object scale variations, further improving overall detection performance. Findings The IMV-YOLO network was tested on the AIR-MAV dataset. Compared to the original YOLOv11, IMV-YOLO achieves improvements of 4.5, 7.7, 6.4 and 6.0% in mAP50, mAP50-95, precision and recall, respectively. IMV-YOLO achieves a detection speed of up to 69 fps, enabling efficient detection of infrared vehicles from multiple angles. Originality/value This paper proposes an infrared multi-angle vehicle real-time detection network called IMV-YOLO. This method can stably detect infrared vehicle objects traveling in different directions under adverse weather conditions, while the detection speed also meets the requirements of real-time detection.