YOLOv5-LC: Enhancing Vehicle Detection for Evening Rushing Hour
Xunming Yuan, Qian Chen, Jinlan Li, Shuyan Gong, Chenxi Lin, Xiaojian Hu · Journal of Transportation Engineering Part A Systems · 2025
In order to address the challenges of low lighting and overlapping vehicles in the detection of vehicles during the evening rush hour, this article proposes a new model YOLOv5-LC based on YOLOv5s. First, to solve the problem caused by the insufficient ambient light in the detection process, CycleGAN is introduced to perform feature domain swapping between the source domain and target domain. The conversion of object features in images is realized by image panning of unpaired images, successfully converting some daytime images in the data set into nighttime images. This enables the model to learn richer feature representations of low-light conditions, thereby improving the model’s ability to detect vehicles under low-light conditions and, in practical applications, reducing the cost of manually annotating nighttime data sets. Second, to improve the model’s ability to detect vehicles under overlapping conditions in the image, the article uses the CrowdDet-V algorithm to supplement the anchor-based object detector. The enhanced detector uses a candidate anchor frame to generate multiple presets to improve the detection ability of the model under the condition of overlapping vehicles in the image, which enables the model to better detect highly overlapping instances in crowded scenes. Finally, the performance of CycleGAN in this task is evaluated qualitatively through image comparisons and quantitatively using grayscale histograms. The proposed model is quantitatively validated on the private COTRS data set for crowded vehicle counting, achieving a 4.3% increase in vehicle detection accuracy compared to YOLOv5s at the cost of little time. The experimental results show that our proposed method can improve the accuracy of congested vehicle detection at night, effectively resolving the challenges of vehicle detection during evening rush hour. On the relatively sparse UA-DETRAC data set, our approach can still achieve moderate improvement, suggesting that the proposed method is robust to different levels of congestion.