Enhancing Nighttime Vehicle Segmentation for Autonomous Driving Based on YOLOv5

Recognizing vehicles in low-light conditions during nighttime poses significant challenges in autonomous driving scenarios due to unclear contours. While instance segmentation models have been extensively studied, their application in autonomous driving night scenes remains relatively unexplored. This paper proposes a method to enhance nighttime vehicle segmentation using instance segmentation models. The BDD100K dataset is leveraged to label autonomous driving daytime scenes and simulate nighttime driving scenarios through data augmentation using gamma correction during the training phase. During the prediction phase, an improved gradient increasment low light enhancement algorithm based on RetinexNet is employed to enhance night driving scene images. Additionally, the proposed method is evaluated using the YOLOv5 model. Experimental results demonstrate that the enhanced YOLOv5 model exhibits significantly improved nighttime segmentation capability, leading to more accurate and robust vehicle segmentation during nighttime. This method shows promise for real-world application in nighttime autonomous driving scenarios.