Improved Nighttime Vehicle Detection Using the Cross-Domain Image Translation

Abstract

Accurate detection of vehicles at nighttime is essential for transportation monitoring and management. However, annotating nighttime vehicle data is challenging, and vehicle features differ significantly between day and night, introducing difficulties in nighttime detection using pretrained models trained on daytime data. In this study, the nighttime vehicle detection performance is improved by employing a patchwise contrastive learning technique to enhance the representation of informative features for various traffic instances. An object detection network with reduced computational complexity and hyperparameters is utilized to conduct vehicle detection at night. Extensive experiments have been performed using images acquired from a section of Jingshi Road in Jinan, China. The impacts of learning rates and crop sizes are discussed. Three commonly adopted indicators, including mean average precision (mAP), precision, and recall, have been used to evaluate the training performance of the adopted FreeAnchor detector. Experimental results indicate that using a crop size of 320 and a learning rate of 2e-4, the developed generative adversarial network (GAN) achieves the best performance in image translation. Moreover, with a ratio of 60% real images to 40% fake images in model training, the FreeAnchor detector achieves the highest mAP of 96.6%. Visualized results for both image translation and nighttime vehicle detection demonstrate improved performance, underscoring the effectiveness of the proposed framework. This study paves the way for leveraging GAN-based networks to assist in vehicle detection under nighttime conditions.