Research On The Identification Method Of Road Markings

Intelligent transportation is the goal of China’s development of smart-network vehicles(smart vehicles,autonomous driving,vehicle-road collaboration).It is also an essential area of development for all the countries in the world.Intelligent transportation can be divided into two aspects: producing very precise maps and building vehicles that can perceive the environment.Therefore,it is vital to create a method for measuring road markings with robust automation and great precision.Due to the variety of types of terrain in China,the vehicle perception system for road markers needs to be different in different cities.Thus,it is also important to build a model operated by vehicle-mounted equipment that applies to other road characteristics,based on the traditional image algorithm for research on road-marking measurement.Simultaneously,based on the theory of depth-wise separable convolutional networks combined with transfer of learning,a study of road-marking detection and road-marking segmentation was conducted.The details are as follows:In this thesis,it is presented that a method of measuring road markings is based on visual sensors.The data on route-measuring road markings were obtained from the images collected by automobile data recorders during the process of driving.First,this method acquires road markings,then corrects image distortion,excluding outside road signs,and restores the road-line parallel relationship.Second,it detects and extracts the road markings,using image segmentation and the sliding window method.Finally,the tracking and measurement of road markers are performed.A technique is proposed that combines a frame-difference method with a statistical model to solve the problem that the characteristics of road markers are very similar and cannot be accurately spliced against the moving background.This method can deal with emergencies(road markers disappearing,increasing,or decreasing,etc.)in the moving stage.Lane-change analysis is used to deal with road inclination caused by cars changing lanes.In this thesis,a road-marking segmentation method based on depthwise separable convolutions is proposed.In the encoding phase,the road-marking feature is relatively regular,so high and low feature fusion is used to extract the traits.This method optimizes the model’s loss function weight by solving the problem of an unbalanced number of pixels between the road and background.It also uses the depthwise separable convolutional layer for feature extraction to reduce the number of model parameters and uses bilinear interpolation for up-sample processing.This process does not generate additional parameters and it can avoid the generation of parameters in the up-sampling process of deconvolution.Simultaneously,this model uses transfer of learning to make it applicable to different environments based on only a small number of samples.The experimental results on road measurement showed that this method has high accuracy and good robustness for road markings from data collected in the laboratory,and it can simultaneously measure multi-lane lines.In the road awareness task,the experimental results in the Tu Simple data set and a self-built library showed that the proposed method can effectively reduce the model’s size compared with other models,and that the model has higher accuracy.The final model can be run on the small NVIDIA Jetson NANO with a processing speed of 23 fps.The two methods that come up within this thesis have value in intelligent traffic applications.