Research On Vehicle Foggy Environmental Perception Algorithm Based On Machine Vision

In recent years,the development of automatic driving technology has become a focus concerned by the whole society.Automatic driving technology is a complex artificial intelligence control technology based on environmental perception,intelligent decision and automatic control.Among them,environmental perception is the most important part and provides a judgment basis for subsequent decision-making and control.With the worsening of the global climate and the increasing of foggy days,it is a difficult problem in the field of automatic driving how to obtain the environmental perception information oriented to the scene intelligent understanding under such weather.The current vehicle driving environmental perception system can be divided into three aspects:weather environmental perception processing,semantic understanding and 3D geometric information.There are still many problems in these aspects,for example,how do you get good quality sharp images,how to balance between processing speed and defogging quality,how to effectively exploit the potential connections in all aspects of environmental perception,and how can we get an environmental perception system that gives consideration to these three aspects.This paper will carry out research on the problems mentioned above.The contributions of this paper are as follows:1)Amended dark channel prior for pixel level is proposedFirstly,we analyze the reasons of halo,failure and color oversaturation for Dark Channel Prior algorithm.Then,the Amended Dark Channel Prior algorithm is proposed to solve these problems.We establish the energy function of dark channel value,in which the deviation of dark channel value is corrected according to pixel point one by one.All problems existing in Dark Channel Prior algorithm can be solved together.Finally,the good subjective and objective processing results are obtained.2)4-directional L1 regularisation algorithm and atmospheric light value algorithm based on Retinex are proposedWe analyzed the advantages and disadvantages of the common filtering algorithms for transmission optimization.According to the unique requirements of transmission optimization,a 4-directional L1 regularisation algorithm is proposed to achieve both texture smoothness and edge preservation.The weight function of L1regularization can be adjusted adaptively according to the specific content of the image,and the image is processed from X,Y,X₄₅ and Y₄₅ directions simultaneously.Compared with other algorithms,the smoothness and edge preservation ability of this algorithm are doubled.We proposed a fast parallel solution of 4-directional L1regularisation filtering algorithm,which can make full use of hardware resources and improve the speed of transmission optimization.An algorithm of atmospheric light values based on Retinex theory was proposed.This algorithm can effectively separate incident light distribution and accurately locate the area with the highest fog concentration in foggy images,so that accurate atmospheric light values can be obtained without the interference of highlighted objects in the image.3)A real-time defoging algorithm for vehicle video is proposedAt present,the defogging algorithm for vehicle video is achieved by simplifying program and taking advantage of hardware resources,these methods improve the speed of defogging while weakening the effect of defogging.So 4DL1R-Net,a high speed defogging network model based on deep learning,is proposed.4DL1R-Net can not only provide high speed transmission optimization results,but also output accurate sky segmentation based on pixel level.So 4DL1R-Net can provide accurate atmospheric light values.It is the first accurate sky segmentation network based on deep learning in the field of image defogging.4DL1R-Net has all the advantages of4-directional L1 regularisation algorithm,and has a processing speed of 0.096 seconds(620×460 pixels).4)A vehicle foggy environmental perception network(FFM-Net)based on multi-feature fusion is proposedThis paper firstly analyzed the potential relationship among weather environment perception,3D geometric information extraction and semantic environment information processing.Then,a deep learning network is proposed,which takes foggy image as input and outputs clear image(transmission map),semantic segmentation and dense depth estimation.And then,we came up with a new concept that the dense depth information of the scene can be uniquely determined by using the foggy image and the corresponding transmission map.Therefore,a deep learning network is used to fit the complex function relation among them,and the dense depth map is finally obtained.5)A traffic scene understanding data set is establishedIn order to train and test 4DL1R-Net and FFM-Net,a traffic scene understanding dataset is established based on Cityscapes,KITTI and many local urban traffic images.This dataset contains 8000 traffic scene images,which can be used for training and quality evaluation of road traffic scene understanding,such as sky segmentation,defogging,transmission and transmission optimization,density depth estimation,semantic segmentation and instance segmentation.