The Research Of Unstructured Road Detection Algorithm Based On Color Features And B-spline And Its Application

Vision-based intelligent navigation system is one of the hot research topics of machine vision, in which road detection is one of the key technologies.The road detection provides accurate position information and direction information for the navigation system.The road detection algorithm must meet the requirements of robustness in real-time.The real-time property means the algorithm must finish the image processing in time and the robustness property means the algorithm can produce the correct output in a variety of complex environment.Currently, as the unstructured roads do not have obvious feature and are coupled with a variety of interfering factors(such as light,shadow, etc.)from the external environment,the detection of unstructured roads is very complex.As the complexity of unstructured road detection,using one single feature-based method or model-based method has been difficult to achieve the desired results.So a good solution for that problem is to combine these two kinds of methods,but how to achieve the trade-off between the real-time and robustness is a difficult problem.To enhance the robustness under real-time requirement is the purpose of our study. Main tasks of this paper are as follows:To sovle the problem of unstructured road detection in non-homogeneous environment,this paper first analyzed the current feature-based road detection algorithm,and laid special stress on the method of Gaussian mixture model(GMM). According to the computing characteristics of GMM Algorithm,a fast block-classify Gaussian mixture color model is proposed:The road image is first processed by block-classify method,then the initial centers of K-means is predicted by kalman filter according to the relationships between two adjacent image frames.Experimental results show that the method can effectively reduce redundant information of the road image and the iterations of K-means algorithm,thereby reducing the complexity of GMM,and inhibiting the noise in the image.In order to improve the validity of unstructured road detection algorithm,integrate the advantages of model-based methods and feature-based methods.This paper proposes an algorithm based on block-classify Gaussian mixture model and B-spline curve.The algorithm first introduces a block-classify method to extract road boundary points set,the mixed area blocks which contain the road boundary are choosen to be processed;A B-spline curve is choosen to be the road model,then the two boundaries of the road are fitted by the B-spline curve in which the best control points were searched by the least square method.Experimental results show that comparing to the max_edge method which use parabolic as road model,our algorithm can flexibility expres roads with various shapes,has a strong anti-interference to external noise and requires little computation.According to the requirements of applications on driveless vehicle,this paper first analyzes the characteristics of driveless vehicle and implements a prototype of the vision-based navigation system,in which the algorithm proposed in this paper was applied and optimized.Lots of field tests proved the navigation system's stability.The algorithm proposed in this paper is robust and real-time in complex environments;it can meet the requirements of intelligent navigation system and have a theoretical value and practical value.