Terrain Recognition With Vision For Mobile Robots

With the development of robot technology and computer vision algorithm theory,the application of robot has been greatly expanded.Outdoor mobile robot should have the ability of autonomous navigation and terrain recognition for normal work.Outdoor natural environment is different from the indoor environment,it is unstructured,complex and changeable.Different terrain will produce different restrictions for the movement of the robot.For the perception of the environment,the vision can provide robot a lot of information.Visual-based terrain recognition helps the robot to understand and make reasonable predictions about the terrain to be passed.Accurate identification of the terrain will ensure that the mobile robot in the natural environment safe and stable movement.This dissertation studies the terrain recognition algorithm based on mobile robot vision.The main work is as follows:Firstly,the research background and significance of terrain recognition are discussed.Recent research of terrain recognition technology of mobile robot is reviewed.The terrain recognition based on vision is compared with that of other sensors.This dissertation analyzes the existing problems in the study of terrain recognition based on vision,and introduces the main research contents and chapter arrangement of this dissertation.Secondly,based on the study of the common data set for the current terrain classification experiment,a more comprehensive SDUterrain dataset is proposed and published.The data set uses the consumer grade camera and webcam for image collection,and the dataset contains light diversity and scene complexity.Thirdly,for the complex and changing characteristics of outdoor scene,this dissertation tried image segmentation on mixed terrain samples,and the watershed algorithm and graph-based theory method are adopted.This dissertation combined the method of graph theory based on comparative theory and segmentation of watershed algorithm,and the segmentation scheme suitable for terrain scene is proposed,which improves the over-segmentation and divides to extract the complete terrain area as well as the regional boundaries,including the horizon.Fourthly,the characteristics of the terrain image are analyzed,and the color histogram,LBP(Local Binary Pattern),SIFT(Scale-Invariant Feature Transform),CEDD(Color and Edge Directivity Descriptor)are extracted from the terrain dataset samples.And this dissertation adopted ELM(Extreme Learning Machine,Extreme Learning Machine)as a terrain recognition classifier.In the experiment stage,firstly the optimal parameter selection experiment of ELM was carried out.After a large number of experiments with different activation functions and hidden elements for each feature,the optimal parameter range is determined.This contributes to the improvement of classification accuracy and saves selection time of hidden neuron numbers.Then,on the basis of this,the terrain classification experiment is carried out on the basis of the single feature.With the problem of insufficient single feature classification accuracy,the scheme of feature fusion is designed and tried,and the color histogram features are fused with LBP features as they have complementary in the way of color and texture.And this fused feature improved the terrain classification accuracy.