Road Traffic Marking Detection And Recognition Based On Multi-level Fusion And Convolutional Neural Networks

In the intelligent navigation system of driverless cars,road traffic marking detection and recognition is one of the key technologies to achieve autonomous driving,which is of great significance to vehicle environment perception and navigation.Because of the existence of non ideal conditions of road traffic marking,for example,long-term exposure to the sun and rain and suffering from vehicle wear lead to fouling,fading,missing and so on,most of the algorithms have poor robustness and low recognition accuracy,and the detection and recognition algorithms of road information are mainly aimed at lane marking and roadside traffic signs,there are few detection and recognition algorithm about road traffic marking.Therefore,aiming at the problem of poor robustness and low accuracy,this paper designs and implements an improved road traffic marking detection and recognition method based on multi-level fusion and convolutional neural networks.Firstly,according to the texture feature and color feature of the image,the method based on vanishing point is used to estimate the road area that can be traveled.Then,using the location of the detected road area,the region of interest in the image is extracted through the saliency detection method of multi-level fusion.Finaly,according to the multi-scale convolutional neural networks of the same layer,the region of interest is identified to determine the type of traffic marking.The main contents of this paper are as follows:(1)In order to improve the detection accuracy of road traffic marking,this paper uses a method based on vanishing point to estimate the road area that can be traveled.With the parameters as one scale and four directions of Gabor transform,the optimal vanishing point of the road is quickly determined by using the local soft voting method by correcting the texture and limiting the selection range of vanishing points.Combining the OCR and color feature of the image,the segmentation of the driving road area can be realized,so as to eliminate the influence of the non-road area.(2)Combined with the relevant saliency area detection algorithms,this paper uses a multi-level fusion method to detect road traffic markings.Using the underlying salient features of the image(color contrast,area contrast,texture contrast),the UCM is fused to obtain saliency map of single layer.The method of linear weighting is used to fuse the saliency maps of different layers to get the optimal saliency map,and the region of interest is extracted by sliding window and non-maximal suppression method.Experiments show that the algorithm can significantly detect the road traffic markings and effectively extract the regions of interest.(3)Using the region of interest detected in Chapter 4,based on the LeNet-5convolutional neural networks,this paper designs a method of road traffic marking recognition based on the multi-scale convolutional neural networks of the same layer.Experiments show that the network model can well recognize 12 types of road traffic markings.Compared with other algorithms,the algorithm has high recognition accuracy(98.4%)and less recognition time(0.011 seconds).