Research On Methods Of Lane Detection In Intelligent Traffic Scenes

Recent years have witnessed the prosperous of vehicle driving assistance systems in intelligent transport community with the help of the rapid advance of computer vision.The exact extraction and detection of lane line,which has been viewed as an important target in traffic scenes,owns great theoretical value and practical significance in the creation of smart cities,the perception of smart traffic environment,and the sharing of information about connected vehicles.Although many efforts have been focused on lane line detection,it is still a challenging task under unfavorable light conditions such as rain,fog and nighttime.Besides,there are some factors that interfere the detection,such as shadow of trees,road stains,and ruined lane line.Most of existing methods concentrate on the fixed scenes with limiting the number of lane line and cause a mismatch in curved road conditions.Aiming at above problems,this thesis studies several technologies involved in vision-based lane line detection based on the previous work,including enhancement pre-processing of lane line detection,using traditional feature extraction and intelligent deep learning methods to detect lane line.The main work is as follows:Firstly,a traffic scene image enhancement method based on contourlet transform and improved Retinex algorithm under weak illumination conditions is studied.The image is transformed into the HSV color space.For night images,the brightness of V-channel needs to be enhanced and then reversed.Then,the V-channel is decomposed into high frequency component and low frequency component through contourlet transform.The high frequency component is processed by threshold filter to enhance the detail of the image and suppress noise.At the same time,the low frequency component is processed by improved an advanced Retinex algorithm to improve the image.The enhanced image is reconstructed from the contourlet inverse transformation.Experimental results show that,this method has more advantages in both subjective visual effect and objective evaluation indexes compared with the four classical image enhancement methods.Secondly,a novel lane line detection method is proposed by combining progressive probabilistic Hough transform(PPHT)and density-based spatial clustering of applications with noise(DBSCAN)clustering algorithm.Based on the perspective characteristics of lane lines,the angle constraint is applied to PPHT to remove the interference of false line segments.Then,DBSCAN clustering algorithm is used to cluster the lane line parameters in Hough space so that the line segments with similar slopes can be clustered together.And the lane line parameters are determined through the core points of the cluster.Comparing with the method using Hough alone,the proposed method is more accurate.The interference and false detected lines are eliminated,and the method does not need to assume the number of lane lines to be detected in the image.Then,in view of that the straight model is not good at approaching curve lane lines,and the curve model has a large amount of calculation and poor real-time performance.To tackle these issues,this thesis develops a model combined a straight model with a curve model.The candidate lane lines in an image is detected by the straight model at first.The vanishing point of the lane lines can be obtained,and the curve of the lane lines is judged according to the relative position of the vanishing point and the candidate pixels.If it is a straight lane line,the improved PPHT and DBSCAN algorithms in the previous chapter perform the detection result directly.If it is a curved lane line,the candidate pixels are fitted by the B-spline curve model through RANSAC algorithm.The experimental results show that,the method performs well on detecting lane lines with different curvature.What's more,the proposed method can handle extreme traffic scenes such as shadow,road stains,ruin lane line and a bad light condition at night.Finally,a lane line detection method based on multi-level and multi-scale feature aggregation of convolution neural network is proposed by combining the features of CNN middle-level and top-level convolution maps to generate lane line feature descriptions.The multi-level feature is obtained by concatenating different levels of convolution output,and the multi-scale feature of the top-level convolution output is obtained by applying several convolutions with different dilation rates.An MFA encoder is constructed.The speed of the model is improved by using the depth separable convolution instead of the standard convolution.The up-sampling decoding is carried out by using the Pixel Shuffle technology.Then the constructed network is trained by the images supervised.The trained model is used to detect lane lines on testing image.The experimental results show that,this method has higher detection accuracy compared with the traditional lane line detection methods and method in paper[48].