Research On Road Detection Algorithm In Complex Weathers

With the rapid development of world economy,autonomous driving has attracted great attention.ADAS(Advanced Driving Assistant System)is the core component of intelligent transportation systems designed to avoid driving accidents,and reduce energy waste and traffic congestion problems.Therefore,many domestic and foreign research institutions have paid attention to road detection in the automatic driving system.However,the current road detection technology based on visual intelligent vehicles is susceptible to external environmental factors,such as complex and variable weathers and the internal factor that is excessive redundant area existing in road,resulting in low road detection accuracy and lacking of real-time performance.This paper takes road detection in ADAS as the research object and conducts in-depth research on the existing problems in the road detection.(1)Research on highlight removal based on threshold filtering and gradient constraint.Aiming at the issues that road color is abnormal,texture is lost and road detection accuracy is affected by traditional algorithm,the dark channel prior and the layer separation method are used to locate and remove the highlight region in this paper.In the highlight region,threshold filtering is employed to locate the strong highlight region that affects road detection.To solve the image quality degradation by preserving the texture and color of the image after highli ght removal,the gradient constraint based layer separation method is used to preserve the texture and color of the highlight layer in the non-highlight layer.Therefore high quality image without highlight is obtained for road detection.The results show that using the proposed road highlight removal method,calculation amount can be reduced and image quality can be maintained,thus the road detection accuracy is improved.(2)Research on defogging algorithm based on dichromatic reflection model and clustering constraint.Due to the insufficient estimation of the atmospheric light coefficient and transmittance,current defogging algorithms always face the challenge of road color distortion which would lead to the failure of the road detection.In response to the problems above,this paper utilizes a dichromatic reflection model to locate and estimate foggy areas.In addition,in order to improve the scene contrast and visibility after defogging,the object function based on clustering constraint is used to remove noise in the fog map.Then,according to the optimized fog map,the local atmospheric light coefficient and transmittance are estimated to reduce the phenomenon that the image color is too dark and texture is lost after defogging,after which the high quality fog-free image is restored.The results show that the defogging algorithm of this paper improves the road detection accuracy and ensures image quality at the same time.(3)Research on rain removal algorithm based on the training on high percentage rain streak image blocks.In view of the issue that the road texture is blurred and the color is distorted which affect the accuracy of road detection in current removal algorithms,this paper uses the binary rain map based on the deep learning to locate the rain streaks and reduce the usage of the rain removal in the non-rain area.In order to improve the ability of the derain learning network,and solve the problem of texture loss after rain removal,the high percentage rain streak image blocks are regarded as training data.And then multi-data fusion and the objective function of gradient constrains are combined to maintain the color and texture features of rain-free image.In this way,clear rain-free images for road detection are obtained.In order to solve the problem of road color distortion caused by the veiling effect,a defogging process is added to the rain removal model for high quality rain-free images and high road detection accuracy.The results show that the proposed rain removal method can reduce the amount of calculation and refine image quality at the same time,thus improving the road detection accuracy.(4)Research on road detection algorithm based on redundant areas removal with vanishing point.Since the accuracy and speed of current road detection methods are affected by non-road areas(sky,trees and high buildings),the real-time performance is poor,which reduces the reliability of ADAS road detection.In order to solve this issue,this paper uses VP(Vanishing Point)to remove redundant areas of images and improve the accuracy and speed of road detection.To speed up the VP detection speed,confidence map and contour map are combined to remove edge interference from non-road areas and reduce VP calculation amount.Then,the non-road area can be removed by using the location information of the VP.Moreover,in order to reduce the network calculation amount,the remaining road area is used to train the road detection network based on CNN.By doing this,the network convergence can be accelerated and the road detection accuracy can be enhanced at the same time.In this paper,to improve the road detection accuracy and speed,the highlight removal,deraining and defogging are used to eliminate the external influence of complex weather on road detection,and the internal influence of redundant area in road image is removed by fast VP.Compared with the most representative methods in this field,the proposed method obtains clearer images without highlight,rain and fog,and achieves higher road detection accuracy and speed.The robustness of road detection in ADAS is enhanced,and these schemes have high practical value in the field of automatic driving.