Towards Extreme Weather Detection: Dark Channel Priori Enhanced Detection with New Labelled Data

Abstract

With more and more cars entering our lives, the number of car accidents is also increasing. To solve this problem, driverless technology has entered our field of vision. An effective way to realize unmanned driving technology is the application of deep learning. At present, the practice of various large companies in this field is mainly to realize the integrated deep learning of image recognition, judgment, and response on automobiles. The model performance is satisfactory in normal condition. However, if the vehicles encounter some extreme weather, e.g., foggy, it is difficult to ensure the safety of deep learning model due to the insufficient accuracy of image recognition. To solve this severe problem, we advance three strategies. Firstly, considering high-capacity backbone will mitigate this situation, Yolov5 with stronger performance was used as the detector in the experiment. Secondly, image pre-process would be benefit to this foggy situation. Thus, we introduce dark channel priori defogging to our detection framework. Thirdly, since deep learning is data hungry, we labelled image collected in heavy fog weather condition and utilize them to train the network. Abundant experiment demonstrates the effectiveness of the proposed strategies. We gain a precision of 0.89734 after training the model 299 times and the safety can be guaranteed with the high accuracy of image recognition.