Research On Surface Cover Classification Algorithm Based On Texture Feature

Vision-based environmental perception system is one of research hotspots in the field of pattern recognition and artificial intelligence, and it is also one of the essential content of intelligent vehicle research. Surface classification is an important part of the content. The study of the structural road is relatively mature which is due to the color, texture, road signs and other features are unified. The current study of the unstructured road in off-road environment is not enough, and the road surface classification result is unsatisfying in bad environment such as under direct sunlight, rain and snow reflection of the sunlight. Intelligent vehicle works in outdoor environment which light, scenery, weather, season and location are complex, unlike in indoor environment which light, scenery are relatively constant. Therefore, this article focuses on the unstructured road surface classification in off-road environment. The main contents of this paper are divided into four parts:(1)We investigate surface texture feature extraction methods, focusing on color feature, LBP feature, SIFT features and feature extraction process. We experiment on the texture database with the combination of different mainstream features and classifiers.(2) We apply SVM to road surface classification. Based on the idea of random forest ensemble learning, KNN-SVM is introduced in road surface recognition. In addition, due to the importance of SVM kernel function, SVM based on hybrid kernel function is also introduced. To classify road surface in off-road environment, one of two-phase methods was adopted, in which the first stage used KNN-SVM joining mixed kernel function to classify grass or non-grass. The experiments proved this way can get higher accuracy rate in this stage.(3)We apply random forest to road surface classification. Based on the idea of maximizing margin SVM, ensemble classification margin and random forest based on classification margin are introduced in road surface recognition. In the second stage of two-phase method, we try to classify dirt road and gravel road (containing 7 categories). The experiments showed random forest based on classification margin can achieve better classification results with less computational time.(4)We investigate the surface classification algorithm based on deep learning. We study the convolution neural network, and present a terrain classification algorithm of unstructured road based on multi-channel convolution neural network. Experimental results show that due to the use of color information, our proposed method has improved greatly with respect to convolution neural network in our database. Off-road road surface classification result shows the method has better adaptability than the traditional method.