| Road traffic accidents have become a main economic and social problem.All kinds of intelligent driving safety assistance systems play an important role in traffic safety.A large proportion of accidents were caused by low visibility of roads in various accidents.So the road visibility detection and low-visibility road information recovery have become one of the key research issues at home and abroad.This Ph.D.dissertation is based on the project of National Key Research and Development Plan named Research on the Key Issues of Sensing Environment,Dynamic Decision-making and Controlling for Intelligent Electric Vehicle(No.2016YFB0100900).In view of the current status of intelligent safety assistance driving technology,this paper takes the Hyundai Motor new Elantra EV pure electric vehicle of the research group as the actual road experimental platform and conducts in-depth research on road visibility detection and road information recovery methods under the unfavorable visual conditions.The different visibility road traffic scenes perception classifiers were designed,road visibility detection method based on road and sky boundary line and lane line vanishing point was proposed,and the foggy road visibility detection method was tested by Hyundai Motor new Elantra EV pure electric vehicle experimental platform.The test results show that the proposed visibility detection method can obtain the foggy road visibility value effectively.Four visibility warning levels were divided to dim foggy road environments and three visibility warning levels were divided to night road environments.Exploring road information recovery method under dim foggy road conditions and image quality enhancement methods under nighttime condition.The research work o this paper specifically includes:Image perception classifiers of different traffic scenes were designed under unfavorable vision conditions based on supervised learning.Firstly,road images with different visibility were obtained from the road video sequences collected by a real vehicle.According to the characteristics of different road images,representative ten road traffic scenes were selected to extract visual features from each image,such as,color features,texture features and edge features.50% of them were selected randomly as input parameters in unfavorable visual traffic scenes image perception classifier training process.By using supervised learning BP neural network,support vector machine,probabilistic neural network,SKohonen network and extreme learning machine for offline learning.BPNN,SVM,PNN,Skohonen and ELM perception classifier models were obtained respectively.Using the other 50% feature set to test the perceptual trained classifiers.It was concluded that the trained extreme learning machine perception classifier was superior other trained classifiers in terms of accuracy and calculation time through the comparative analysis of accuracy and running time.Foggy road visibility were detected based on the boundary between sky and road and lane line vanishing point.Firstly,the Hough transform straight line model was used to identify the left and right straight lane lines.The intersection point was obtained based on the lane straight line model.The intersection point of the lane line is the vanishing point of the foggy road.The image coordinate value of the road vanishing point was estimated by using the transition relationship of the Hough transform polar coordinate system to the Cartesian coordinate system.Then the region growing segmentation method was used to obtain the boundary of foggy sky and road.According to the image characteristics of foggy roads,the visibility of foggy roads depends on the relative sky height.Therefore,this paper analyzes the geometric projection relationship between the image coordinate system and the world coordinate system.Finally,the foggy road visibility calculation model was derived by using the y coordinate axis of the vertical direction of the image physical coordinate system.A method to classify the foggy road visibility level was proposed based on supervised learning.Firstly,the fog image features that can represent different degrees of fog concentration were analyze,including average gradient,entropy,contrast,edge intensity,rate of dark channel pixels,and transmittance features.The above six characteristic parameters were used as the input parameters of the visibility warning level recognizer.And the support vector machine and the extreme learning machine visibility warning level recognizer were designed respectively to be used to classify the foggy visibility warning level under different foggy conditions.A night visibility detection method was proposed based on night image features.Firstly,the formula for calculating the nighttime visibility of point sources was derived based on the Ollard's law formula.The parameters include the original intensity of the target light,the different nighttime test distance,and the illuminance that was seen by the observer.In order to obtain the illumination of the nighttime source,the nighttime light sources were obtained through the balser industrial camera.Then obtained original images were respectively processed by image size normalization,an adaptive threshold segmentation,source region localization.The total gray values of the light source were obtained by the VC++6.0 MFC dialogue application program,and the nighttime visibility value is calculated by the deduced the point light source visibility calculation model.The test results show that the visibility values obtained by the proposed method were in accordance with the law of atmospheric extinction.A night visibility warning level classification model based on supervised learning was established.Firstly,in order to obtain image features that represent nighttime roads scenes of different visibility,a large number of nighttime road images were analyzed,including average gradient,contrast,edge intensity,brightness spatial component of HSV,and Fourier transform spectrum amplitude.The 50% of the above features were used as input parameters of the nighttime visibility warning level recognizer,different night visibility road scene warning level recognizers were designed based on SVM and ELM supervised learning models to classify different visibility night conditions.The methods of nighttime image enhancement and the foggy road image information restorationwere proposed.Firstly,a SSR-DCP method based on fusion of single-scale retinex(SSR)and dark channel prior(DCP)de-fogging was proposed to realize the image restoration of dense foggy road scenes.It has been verified that the proposed SSR-DCP method can improve the halo phenomenon that processed using only the DCP method.According to the characteristics of night road image,the improved SSR algorithm was used to restore the nighttime low visibility environment image,and the enhanced image can achieve daytime visual effects.From the objective and subjective evaluation comparison analysis of the McCann99 Retinex,Frankle-McCann Retinex and improved SSR methods,the experiment shows that the night image visual effect processed by improved SSR algorithm are more meet human visual characteristics.The improved SSR method is better than the other two methods in terms of the time complexity of image information recovery.Foggy road visibility detection of real vehicles road test was designed.The Hyundai Motor new Elantra EV pure electric vehicle of the research group was used as the test platform.The balser industrial camera was installed on the windshield of the interior rearview mirror,and the MFC dialogue application program based on lane departure warning and road visual visibility detection system was written by VC++6.0 programming environment.The road real vehicle verification test was performed by a road visibility detection system based on road and sky boundary line and road vanishing point.The test results show that the visibility detection method can obtain the fog road visibility value accurately.In summary,the paper proposed environment perception classifier model under the unfavorable vision,road visibility detection method based on road and sky boundary line and lane line vanishing point,the foggy road visibility warning degree classification model,road information recovery method of the dark foggy,the night road visibility warning level division method and the night low visibility image information recovery method,all of them can be used as important safety countermeasures to overcome threats to driving safety under the adverse visual road conditions.The proposed environment perception and road information recovery methods in this paper can be applied individually or in combination.It is of great significance for the research of unmanned vehicles road environment perception technology and vision-based automotive intelligent safety assistance system. |
PDF Full Text Request