Unstructured Road Detection Based On Monocular Vision

With the development of science and technology,more and more countries are carrying out research on driverless technology.Road environment detection,as the basis of autonomous driving technology,directly affects the performance of autonomous vehicles.Visual navigation technology has the characteristics of simplicity,flexibility and low cost.Roads can be divided into structured and unstructured roads.Structured roads have clearly marked lane lines by hand;Unstructured road has no manual marking,and often the structure is changeable,there are shadow,water and other factors.This paper aims to realize more accurate detection of the unstructured road with interference and the recognition of obstacles on the road under the moving state.The road detection is based on the information entropy based image segmentation and optimized.Firstly,image preprocessing is used to eliminate noise interference,compress data volume and enhance detailed information.Secondly,several image segmentation methods are analyzed and compared,and the pixel is preliminarily qualitative by using the fast recursive quadratic two-dimensional maximum entropy segmentation algorithm.An adaptive template was introduced to calculate the Mahalanobis distance from the point of the undetermined area to the road area and non-road area,extract features,and input them into the SVM classifier to optimize the segmentation effect.Finally,considering the structural characteristics of the boundary region,the improved segmentation algorithm is used to segment the image into blocks for fast extraction of edge points.The random sampling consensus algorithm is used to further optimize the edge points and eliminate the interference of pseudo-edges.The quadratic curve model is selected and the road boundary is fitted by the least square method.Through the simulation experiment,for the three unstructured road scenes,the average accuracy is 80.4%,which indicates that the improved algorithm can identify the unstructured road more accurately,and has a certain robustness against the common shadow,water and other kinds of interference.The recognition of obstacles is based on the optical flow method and optimized.Firstly,according to the characteristics of obstacles under the road scene,the most appropriate optical flow method is used for obstacle identification.Secondly,the background vector estimation and interference information filtering are performed by combining the constraints of road area.Finally,significant differences were used to enhance the discrimination between target and background.Simulation results show that the algorithm can detect and recognize moving targets in many scenarios.