Research on sensor dirt recognition of autonomous vehicle camera based on deep learning

Abstract

Intelligent driving technology is an important direction of the current development of automobile technology. The realization of intelligent technology is based on the perception ability of car sensors such as cameras and lidar for the driving environment detection. In the actual driving, the surface pollution of sensors will reduce the safety of intelligent driving vehicles. Therefore, it is very important to realize the automatic recognition of ultra-near-field dirt on the optical surface of car sensors. At present, the research on camera sensor fouling recognition is very scarce, and the relevant public data set for recognition research is also lacking. This paper conducted an in-depth study on the contamination problem of the camera sensor of the autonomous vehicle, independently created a dataset of dirt on the camera sensor, and trained the dataset based on the YOLO-V5 model using this dataset. The results show that the dataset we constructed is quite reasonable, with obvious features and excellent training results, with mAP(0.5) reaching 0.987; The results of the detection experiment prove that the trained dirt recognition model has certain universality, and has relatively excellent detection effect on different datasets, and the accuracy rate and recall rate can reach more than 0.8. This work could provide a certain reference for the cleaning research of automatic driving camera sensor.