Object detection under bad lighting condition for autonomous vehicles for rain images

Abstract

Machine vision is only a part of auto-driving sensing system, but it is the most basic and critical part. It is to detect vehicles and traffic signs/lights. And object detection algorithm plays a critical role in the machine vision. With the development of science and technology, target recognition has developed from the initial manual method to computer automatic recognition algorithm, which greatly improves the accuracy and efficiency of recognition. Because the specific environment and interference of target recognition are very complex, there is still no general algorithm suitable for many environments. In this research project, we first made a picture data set in rainy environment for auto-driving vehicles research. And we learned the adversarial generation network technology. Based on an open-source KITTI dataset, we used the CycleGan network to generate a large-scale dataset of autonomous vehicles in a simulated rainy environment. After that, we extensively investigated different algorithms in the field of target recognition, including different representative algorithms of one-stage and two-stage. The evolution history and route of technology in this field are understood, and several algorithms are tested, and the different performances of different models in rainy environment are obtained.