Low-Light Animal Detection on Highways Using Enhanced YOLOv5 and Image Preprocessing Techniques
Year: 2026
Type: conference-paper
Abstract: Collisions between animals and vehicles on roadways are an ongoing topic of safety and conversation concern to humans, particularly when the visibility is severely limited, such as at night time. The traditional monitoring and detection systems have low illumination, glare, and environmental noise that can contribute to the late detection of a situation that results to accidents. To address this problem, we gathered a selected set of images of 927 images by Roboflow that had been captured in night forest and highway setting and annotated in three large animal groups: deer, fox, bear. The proposed technique was the combination of Contrast Limited Adaptive Histogram Equalization (CLAHE) and a Robust Retinex Model to perform illumination correction, YOLOv5 to detect objects in real-time. The hybrid front end offers the advantage of having a better visibility whilst the reduction of false positives caused by glare provides a stable detectable performance in low light. An additional feature of the work is its two-stage pre-processing-detection fusion, consisting of a illumination normalization step and light-weight high-speed model suitable for real-time inference. Model comparison with Precision, Recall, F1 Score and mean Average Precision (mAP) showed the great improvement against naive detectors. Our improved YOLOv5 showed Precision$=0.923$, Recall$=0.773$, mAP${@} 0.5=0.802$, and it achieved 55-60 FPS processing frequency, demonstrating that the detector was efficient and reliable for real-time night highway safety operation.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202219981