Terrain Classification And Natural Scene Segmentationin Field Robot

Computer vision is always the one of most essential subjects in artificial intelligence.In robotics especially in field robotic design,the technology of terrain classification and scene segmentation becomes the foundational structure about movement adjustment,navigation,information collection around computer vision.According to multi-dimension information,correct feature model,large data base,terrain classification and scene segmentation system provides a accurate classification result.Because of the too large,non-isolated data with uncertain disturbance,higher stability and accuracy are required.A set of terrain resource are proposed and used in subsequent work.By analyzing traditional terrain classification and scene segmentation methods with strengths and weaknesses,propose a way focus on terrain classification and natrual scene segmentation,that meet the requirement of classification and segmentation accuracy and garantee real-time performance.Firstly,based on the meaning of terrain classification and scene segmentatio and the research status,methods and their development are discussed.Otherwise,various methods and their limits are proposed,analysis of the inner cause is for exploring the better way to terrain classification and scene segmentation.Standard image devices,construction of data set SDU-Terrain and programming of image processing platform SDU-TC are described in detail.The establishment of standard set benefits the maintainance,unity of data and classification effect in final.And RGB color feature and implement are introduced including RGB histogram feature and square deviation curve feature.For higher classification curracy and solution of original defect,rotational and symmetric invariant LBP is introduced and its principle and advantage are mentioned.Through improvement of several methods of terrain classification and scene segmentation,several methods of different feature fusion and several classifier based on different type feature are proposed.Finally,based on the preceding part about terrain classification,for field robot vision control system.a fusion feature watershed scene segmentation method is proposed.It is a improved watershed segmentation method.and its work principle and iteration calculation process are mentioned in detail.Image segmentation accuracy measurement is always a difficult problem.A objective accuracy meassurement method about terrain scene segmentation is proposed to provide a constructive feedback for whole system design.