YOLO-AOD: A lightweight algorithm for detecting foreign object invasions on railway tracks in foggy conditions

Considering the performance degradation of object detection algorithms in foggy conditions, a lightweight end-to-end detection algorithm is proposed by adding a dehazing module AOD-Net to the front end of YOLOv8 model, called YOLO-AOD. The architecture of the proposed algorithm is composed of two modules: a dehazing module and a detection module. Firstly, the dehazing module employs a lightweight dehazing network, AOD-Net, which is to enhance image clarity and reduce the interference of fog on detection performance. Secondly, the detection module employs YOLOv8n algorithm as a baseline framework, with targeted modifications made to accommodate the characteristics of dehazed images. Thirdly, an enhanced FasterNet block is utilized in the construction of C2f-RGN-F module, thereby achieving model lightweighting and enhancing feature expression capability. Furthermore, the shallow feature information from backbone network is fused in neck network, thereby effectively addressing the issue of detail loss caused by multiple convolution operations. This enhancement of the model’s ability to detect distant foreign objects (small targets) is achieved. Finally, by combining CIoU and normalized Wasserstein distance, a new N-CIoU loss function is designed to improve the model’s robustness in locating boundary-blurred targets. The experimental findings demonstrate that on a self-built foggy track foreign object dataset, YOLO-AOD attains an mAP@0.5 of 80.4%, an inference speed of 94 frames per second, with a concomitant reduction in parameters and computation to 2.7 M and 7.3 GFLOPS, respectively. This significant performance enhancement over YOLOv8n and various mainstream detection models is a notable achievement. The findings substantiate the effectiveness and practicality of this approach in foggy conditions.