Vehicle Detection at Night Time

Recent growth in deep learning has opened up many opportunities for the problem of vehicle detection. Detecting objects in poor visibility is catching scientists attention. In this study, we choose night as the challenge. We conducted training and evaluation of the YOLOv4 method in combination with image preprocessing methods: gamma, CycleGAN's night-day conversion model was retrained on DETRAC data. Night dataset (26,168 images) extracted from DETRAC were used. The results showed that the training on the primitive data is highly effective (64.51%mAP) compared to the image changed from night to day, particularly on the car class (92%AP), bus (91%AP). This is the premise for the next studies and the basis to develop intelligent traffic monitoring systems.