DR-YOLO: Research on Low Illumination Track Foreign Object Intrusion Detection Algorithm Based on YOLOv5

In response to the issues of poor clarity, color loss, and lack of texture details in railway traffic images under low-light environment at night, which leads to detection accuracy, an algorithm DR-YOLO is proposed for the detection of foreign object intrusion on railroad tracks under low-light environment. Firstly, a feature extraction module incorporating depthwise separable convolution is proposed to enhance the model's ability to detect objects of different sizes in low-light environments. Secondly, by utilizing a multi-scale feature fusion pyramid network structure, the effective information in low illumination image features is enhanced, capturing potential correlations between samples. Finally, an attention mechanism module is introduced to obtain contextual and local information of features in the spatial domain, thereby enhancing target features in low light images and effectively improving model detection accuracy. The accuracy and effectiveness of the foreign object intrusion detection algorithm were validated through comparative experiments. The experimental results show that the improved algorithm increased mAP50 by 4.6 percentage points on a self-constructed night track foreign object dataset, with a maximum mAP50 of 76.8%, Compared with the unimproved algorithm, it effectively improves the accuracy of foreign object recognition on low light trajectories, proving the effectiveness of the new algorithm.