Robustness Evaluation of Rail Track Detection Models Under Natural and Adversarial Disturbances

The input data used in track detection models are significantly affected by natural environmental factors such as rain, snow and fog, which can compromise detection reliability. Additionally, deep learning-based detection networks are inherently vulnerable to adversarial attacks, which can lead to incorrect predictions and result in issues such as missed detections or false positives. To address these challenges, this paper simulates various natural disturbances through image enhancement techniques and proposes two black-box adversarial attack algorithms - StrRS and DAG-NES - targeted at object detection models. This study examines how the detection model's performance deteriorates when exposed to both natural environmental disturbances and adversarial perturbations. Experimental results demonstrate that, under the joint effect of natural disturbances and StrRS, the model's accuracy drops by an average of 74.34 %, recall by 90.18 %, mAP@50 by 87.20%, and mAP@50-95 by 88.81%. Under the combined influence of natural disturbances and DAG-NES, the average reductions in accuracy, recall, mAP @ 50, and mAP @ 50-95 are 38.72 \%, 61.90 \%, 54.43 \%, and 62.96 %, respectively.