The Research On The Method Of Foggy Image Simulation In Road Scene

Foggy days have varying degrees of impact on people’s travel and daily life,and are one of the important causes of traffic accidents.Object detection based on road traffic images is one of the important components of intelligent transportation systems and one of the core research contents of autonomous driving technology.However,foggy weather brings great difficulties to the detection and recognition of target images.For example,many traffic sign recognition technologies cannot meet the needs of practical applications under low visibility.In response to this problem,the thesis analyzed the foggy image,and explored the fog simulation and application method of the image in the road environment according to the characteristics of the foggy image.It is of scientific research and practical application significance to solve the problem of the lack of traffic image data in fog and to further study the recognition of fog traffic image.In order to better simulate the foggy weather on the road image,the thesis firstly analyzes the principle of fog formation,the reasons and characteristics of foggy image degradation.There are a large number of unevenly distributed suspended particles of different sizes in the foggy environment,which have an attenuation effect on the propagation of light.The thesis compares the difference between clear outdoor images and foggy images in the same scene,and determine the main characteristic evaluation indicators for foggy image quality.The thesis uses four image quality evaluation indicators related to the characteristics of foggy images,including gradient function,variance function,gray difference function,and information entropy function,to evaluate the characteristics of foggy images.Secondly,based on the scattering theory of atmospheric light and the a priori theory of dark channel,the thesis solves the atmospheric light value and transmittance value of the foggy image.The thesis established a foggy image simulation method and a visibility estimation method.In the process of research,it was found that the estimated value of transmittance has a large deviation,which brings the problem of halo effect to the image.The thesis therefore uses guided filtering to refine the transmittance.Finally,based on the fog image simulation method,the thesis conducts an image fog simulation experiment on the LISA data set containing traffic signs.Built image datasets of various foggy traffic signs.Based on this data set,the thesis uses the deep learning network model YOLOv4 for training,detection and recognition,which solved the problem of insufficient traffic sign images in foggy weather in road scene.The foggy day image collection established based on the simulation method of the thesis was used for specific application simulation.The thesis analyzes and compares the simulated foggy image and the original foggy image.The results of the foggy image simulation experiment show that the error value of each image evaluation index of the simulated foggy image is less than 5%compared with the original image,which verifies the reliability of the method.The thesis trains the YOLOv4 network model to detect and recognize foggy traffic signs.The accuracy of the YOLOv4 network model on the traffic sign recognition in the test set is 88.7%.The model achieves a good recognition rate of traffic signs in fog.This model can perform actual foggy traffic sign recognition,which basically meets the needs of foggy traffic sign detection.It further verifies the feasibility of the foggy image simulation method proposed in the thesis,and the established foggy traffic sign image data set can be used for more in-depth foggy image research.