Research Of The Unstructured Road Detection Based On Machine Vision

With the development of artificial intelligence,more and more companies are working on the technology for autonomous vehicles.At present,the theory research on structured road detection is quite mature,and autonomous cars also choose the highway or urban road with high degree of structures on the test.But the study of unstructured road is still in the laboratory stage now,due to the complexity of the environment.Intelligent driving consists of three parts: road recognition,obstacle detection and traffic sign detection.In this paper,aiming at the road recognition,a method,unstructured road detection based on machine vision,was proposed.This method is mainly divided into three parts: the static image preprocessing,segmentation,and the edge recognition.In part I: The RGB image was converted to HSI color space firstly.In HSI space,the color information only contains in hue and saturation components.Then,using the median filter algorithm removes the image noise.To predict the location of the finishing line in hue component,this paper uses the normalized cross-correlation algorithm.According to the location of the finishing line in hue component,the hue and saturation components are saved as a sub-image with the same size.In part II: The method,2-D minimum Tsallis cross entropy,are used to calculate the optimal threshold in hue and saturation components.Compared with the Shannon entropy,the Tsallis cross entropy contains a cross correlation coefficient,so it can improve the accuracy of the threshold segmentation.Through the analysis of the binary image after segmentation,this paper adopts the logical operations to reduce the fusion image's dimension.In the fused binary image,the road area is mostly white,while the non-road area is mostly black.And compared with 1-D OTSU,2-D OTSU in gray images and 2-D OTSU in color image,the method this paper proposed is the best.In part III: To improve the quality of binary image,using morphological operations smooth contour in fused binary image firstly.According to the principle of curve fitting,the fused road image can be divided into several segments.In each segment,it recognized the unstructured road edge via the improved double dogleg method.In the final image,the white area is passable.Compared with the commonly Hough line test,this method is more universal.Choosing roads with different materials,such as asphalt,cement,dirt,as well as shadow and bend conditions for more experiments,the results show that the method in this paper is effective.In this paper,all algorithms codes and PC interface are written with MATLAB language,.Different from the traditional gray image segmentation,this paper uses color information components to segment directly.To reduce the amount of entropy calculation,the normalized cross-correlation method is introduced to predict the finishing line location in hue component.In this way,it can reduce the image's size significantly.