Trapezoidal Model And Support Vector Machine-based Unstructured Road Detection

The autonomous navigation system of intelligence vehicles is an integrated system including video capture, road detection, auxiliary driving and other technologies. Road detection is one of the key technologies of the autonomous vehicle navigation system, and the frames captured from the video are important sources of road detection in sensing the environment. In this paper, the road is detected for the video capturing in the road.The road can be classified to structured road and unstructured road. The structured road has obvious lane lines, lane markings or road boundaries easily recognized; while the unstructured road has no or indistinct boundaries and the lane boundaries are relatively hard to be recognized. There are only a few methods to research on the unstructured road, and this paper is road detection of the unstructured road. As the support vector machine (SVM) is the best classifier in the theory, this paper proposed a method using support vector machines to do road detection based on Trapezoidal model and BP(Back-Propagation) network(proposed by H. Jeong, et al).There are several main components:1.A lane detection approach based on trapezoidal model and SVM is proposed. The frames extracted from the video are pretreated by PCNN (Pulse Coupled Neural Network), and then processed by Kalman filter and EM (Expectation Maximization) algorithm. Next, using SVM gets the result of the lane detection and using morphological filter and edge extraction get the final detecting result. This method achieved better classification than the method proposed by H. Jeong, et al.2.For the different features of the road,a model-oriented lane detection approach using fuzzy SVM is proposed, further simply dividing the unstructured road into straight road and curved road,so these two corresponding SVMs are designed to process these two types' road. We use the fuzzy algorithm to judge which SVM would be selected. Next, do the further lane detection, and then using morphological filters and edge extraction obtain the final detecting result. This method achieved better classification than the first method, and better generalization in these roads of different types.3.The lane detection approach using FSVM method exists to determine the error which easily affecting the SVM classification results and contains limited different types of roads, a new method using the Ada-SVM to do the lane detection is proposed. Ada-SVM is developed, where a set of moderately accurate RBFSVM is trained for AdaBoost by adaptively adjusting the values of a instead of using a fixed one. This gives rise to a successful SVM based AdaBoost. Then form the final classifier to detect the lane boundaries. This method had the better classification and higher generalization than H. Jeong's method and these two road detection methods ahead.