Cross Different Time Periods Detection Algorithm based on YOLOv4

With the development of autonomous driving vehicle target detectiontechnology, multi-target detectionincomplex environment has gradually become the focus of attention. In this paper, based on the YOLOv4 algorithm, the target detection problem is carried out on the images from different time periods in the training and test set. Firstly, we train the unsupervised CycleGAN model to get some synthetic images at night. Secondly, these synthetic night images combined with the images collected in the sunny weather are input to train the YOLOv4 detector and then the final model is used to detect the targets at night. The experimental results show that the detection accuracy of our model for car, person and traffic sign is improved by 3.65%, 4.29% and 5.42% respectively when compared to the baseline model, that is, only training the images from the daytime and testing the targets during the night. Inaddition, the recall values of these three targets are all increased by more than 5%, which indicates that our model reduces the missing detectionrate. The above experimental results have fully demonstrated the effectiveness of the combination of supervised and unsupervised algorithm for target detectioncross different time periods. And the proposed model greatly improves the safety of the autonomous driving vehicles, which is of great significance for target detection in the road.