A Detection And Identification Method Of Road Traffic Markings Combining The Visual Perception And Learning Approach

In China, there are about more than two hundred thousand people died in the car accident each year. Up to 90% of these accidents are caused by faults of drivers [1]. Autopilot helps reduce accidents due to human factors, and right driving according to the law of road traffic in the structured road is a challenge in the automated driving technology. The premise of right driving in accordance with the laws of traffic is the detection of road markings, such as: how to detect straight instruction markings and turning markings, the lane markings and no parking instruction markings, etc. In this Paper, we take the road traffic marking as object of study, at the same time, we analyze the lower detection accuracy and poorer robustness on the existing detection and identification method based on the model, visual attention mechanism and the learning mechanism. What’s more, traditional road detection mainly aims at the detection of lane and roadside traffic signs, while detection and recognition for road traffic line markings is rare.Based on this, the paper first analyses road traffic line markings and saliency correlation principle. Secondly, on the basis of utilizing various features of traffic markings, we put forward road traffic line detection method based on visual attention mechanism to detect the road traffic marking. In view of the unbalanced weight distribution of the sample and the abnormal value of the training set, this paper put forward an identification method of road traffic markings combining the visual perception and learning approach for identification. Finally, under the structured road environment, the paper tests and verifies the algorithm based on this data set. The research content and main innovations of the paper include the following aspects:(1) Present a road traffic line detection method based on mechanism of visual attention.Firstly, on the basis of color feature vector and contour feature vector, we extract space context feature, and establish a road traffic marking significant test model based on hierarchical context information. Then, we adopt the method of cosine similarity measure to fuse several layers of significant features, through this we get the road traffic marking significant figure. Finally, by using the method of sliding window and the maximum inhibition to locate of road traffic markings coarsely.(2) Put forward an identification method of road traffic markings combining the visual perception and learning approach.On the analysis of the multi-classification Real AdaBoost algorithm, this paper proposes an outliers-robust MR_AdaBoost algorithm, mainly aiming at the problem of unbalanced weight distribution of the sample and the abnormal value of the training set. Through the classifier, the test image can complete the identification of road traffic marking.(3) The testing and validation of algorithm under the environment of structured road.Results show that compared with mainstream saliency detection algorithms in the dataset of this paper, the suggested algorithm of combing the visual perception and learning to detect the road traffic marking has better performance in MAE and PR curve. Algorithm of combing the visual perception and learning to recognize the road traffic marking can have more than 90% of the precision and has high true positive rate and low false positive rate in multi-classification.