Segmentation Research Based On SVM Road Scene

Ground intelligent robot is out of the direct control of human, and can be autonomous controlled. The major concern in the research of ground intelligent robot is autonomous navigation, and the key technology of autonomous navigation is the navigation scene segmentation. Ground intelligent robot understands the working environment mainly based on visual information, through visual navigation sensors getting road scene images, segmenting the road scene images, and identifies road area. In the outdoor unstructured environment, because of the complexity of the environment factors, such as weather changes, light intensity and other factors, it brings difficulties to image segmentation, and is not conducive to intelligent robot autonomous navigation. How to extract the features of stability and high distinguish to describe the complicated and changeable environment, how to split the complex scene image efficiently, is the main content of the research on ground intelligent robot autonomous navigation.The focus of this paper is researching the road scene image segmentation in ground intelligent robot autonomous navigation, and the main work is as follows:(1)This paper studies the principle of road scene image segmentation, support vector machine (SVM) method and its related theory, and studied the road scene image segmentation method based on b-BTSVM.(2)This paper studies the method of shadows road scene segmentation using b-BTSVM based on super-pixel. Segmenting the image into a lot of super pixel where the pixels have the characteristics of homogeneity and the same lighting conditions by using the SLIC Super-pixel (simple linear iterative clustering Super-pixel) algorithm the k-means clustering algorithm which based on pixel color information and position relationship. This paper put forward a method of feature window orientation, searching for the window which is fit for super pixels region to extract features. Our method calculates the texture characteristics of the super pixels by LBP descriptor which is robust to illumination change, and combines with HSV color space characteristics which are not sensitive with light into a high dimensional feature vector. Learn all kinds of targets in images by b-BTSVM, and establishing a classification decision. In the test phase, segment the scene image into a lot of super pixels by the same method,of which extract the feature vector, which is tested by the existing classification decision surface, to recognize weather the super pixels is belong to road regions. (3)This paper studies the method of based on Bayesian based binary tree support vector machine (b-BTSVM) road region online segmentation. First, initialize the first road scene image by K-means method, and choose certain samples from the district domain respectively as training set; second, use b-BTSVM to learn training set to calculate the classification decision, for the next road image detection; third, update the training set to calculate the classification decision by online learning method, so do detect again.