Feature Fusion And Semantic Segmentation For Terrain Recognition

Terrain Recognition technology as an important research topic in the field of machine vision,which is widely used in mobile robot research.Terrain Recognition is the premise of localization and navigation for autonomous mobile robot.It can help robot understand the surrounding environment,determine terrain types so that the robot can adjust control decisions and path planning.Compared with the wheeled robot,the foot-type robot has a lower speed and higher energy consumption,but it has better trafficability and is more adaptable in unstructured environment.Therefore,it has been widely concerned by scholars at home and abroad.Compared with the method of terrain recognition using laser and other sensors to obtain information,images are convenient and can provide abundant environmental information.Combined with machine learning algorithm,it is more advantageous for mobile robots to traverse unknown and complex environments.The main work is as follows.Firstly,this dissertation discusses the application background and research significance of terrain recognition,summarizes and analyzes the current research status and existing problems at home and abroad.Second,there are few datasets in the direction of the terrain recognition at this stage.In this dissertation we establish the image datasets contains six typical flat terrain.Gather the information of the environment by monocular vision,such as cameras,mobile phones under the different light and weather conditions.Thirdly,feature extraction algorithm based on visual image is studied.Aiming at the problem of classification of flat terrain,this dissertation starts with the traditional machine learning,and mainly innovates in the selection of terrain features.The low-level features based on vision can be targeted to express the characteristics of the image.But the information is incomplete.Deep feature is extracted by self-learning from neural network,and contained abundant information.There is a semantic complementary relationship between the above two,which effectively improves the classification performance in terrain recognition research.Fourth,the problem of terrain segmentation and classification in complex scene is solved.Camera usually captured images contain multiple types,the segmentation and recognition needs to be done at the same time.The semantic segmentation can realize the combination of segmentation and recognition.DeepLab is selected to segment terrain in scene by combining with low-level feature map in input images.In this dissertation,this method achieves good segmentation effect in the reprocessed Sift-flow dataset.Finally,it summarizes all the work of this dissertation,points out the unsolved problems in terrain recognition and looks forward to the next research direction.