Light Enhancement Algorithm Optimization for Autonomous Driving Vision in Night Scenes based on YOLACT++

Abstract

In scenes with low lighting at night, the outline of objects that need to be recognized, such as vehicles, is not clear, and cannot be accurately recognized by the automatic driving system. At present, there are many researches on instance segmentation models, but there are few researches on the instance segmentation application of automatic driving night scenes. According to BDD100K dataset, the automatic driving daytime scene dataset is marked. First, we perform data augmentation by using gamma correction to simulate the night driving scene in the training phase. Then we use our improved low-light enhancement algorithm with gradient increment based on RetinexNet in the prediction phase to brighten night driving scene images. Furthermore, we evaluated our proposed method on YOLACT++ model. The results show that the improved YOLACT++ automatic driving night segmentation ability has been significantly improved, the segmentation of vehicles at night is more accurate and robust, and it has better application value in night automatic driving scenarios.