Design and Evaluation of a Vehicle Detection System in Low Light Conditions

The advent of autonomous vehicles has motivated researchers to focus their research on developing improved advanced driver assistance systems (ADAS). A key area of study in ADAS is vehicle detection. Many have proposed vehicle detection systems in daylight conditions but very few have proposed vehicle detection systems in night time. We have proposed a vehicle detection system in low light conditions in this paper. Cycle Generative Adversarial Networks (CycleGAN) is used to translate images from night to day. This makes use of unpaired imageto-image translation. Super Resolution Convolutional Neural Networks (SRCNN) is used to improve the resolution of the image. Finally, several object detector models like YOLOv4, Faster-RCNN, SSD, ResNet are used to detect vehicles. The results are compared. We have achieved a maximum mAP of 61.9% using the proposed model.