| With the development of computer vision technology,road detection based on vision is playing a dominant role in the navigation system of unmanned vehicles.Because of the complexity of the environment and the uncertainty of the road conditions,the visionbased unstructured road detection is a challenging task.Only the real-time and accurate detection of the driveable area of the unstructured road can realize the true driverless driving.Therefore,it is of great scientific research value and practical significance to detect and identify the driveable area of unstructured roads.The paper will study the methods of detecting unstructured roads.The main research contents are as follows.(1)Based on the texture method,the research of unstructured road vanishing point detection is carried out.Aiming at the lack of unstructured road edge information and rich texture features,the paper uses texture-based methods to detect road vanishing points.Gabor filter is used to extract and analyze the texture features of road image.For the approximate consistency of road texture direction,the method based on confidence interval is used to filter the effective voting points.According to the relationship between vanishing points and voting points,the soft voting method based on distance weighting is used to vote on vanishing points,thus realizing the detection of road vanishing points.(2)Research on unstructured road detection method based on region growth and wavelet transform.In the paper,the region growth method is used to segment the unstructured road initially,and the growth seed points and growth criteria are effectively designed according to the location of the road region in the road image.In view of the under-segmentation or over-segmentation of the road edge by the traditional segmentation method,the maximum modulus of wavelet transform is used to detect the edge of the road image,and the obtained edge information is used to modify the region growth method to initially segment the road image to obtain the road segmentation image with complete edges.The detection of unstructured road lane line is realized through road area extraction,road boundary point extraction,road boundary point fitting and other operations.(3)Research on unstructured road detection method based on Mask R-CNN.The paper selects 1000 unstructured road images,and finally collects more than 1500 road images through data enhancement.In view of the poor segmentation effect of traditional segmentation algorithms in complex environments,this paper studies the theory and structure of the Mask R-CNN instance segmentation algorithm,uses unstructured road data sets to train and test the Mask R-CNN model,and finally the m AP reaches 0.84.The Mask R-CNN model is used to perform edge extraction,edge radial detection,curve fitting and other operations on the mask image predicted from the test image to achieve the detection of unstructured road lane lines. |
