Research On Online Terrain Classification Methods Based On Vision And Vibration Information Fusion

With the application of tracked mobile robots in the unstructured environmentdeveloping, demands for ability of robots to perceive the environment, especially forability to sense the complicated ground are rising. Complex terrain can be described bygeometric parameters and physical parameters. Study of physical parameters whicheffect the performance of robot’s continuous operation and self-stability become thehotpot. Focused on the issues related to physical parameters and aiming at characteristicof the terrain type, this paper study on the method of online terrain classificationmethods based on vision and vibration.Combined with characteristics of visual signal of the complicated ground typeprocessing for the captured image. Basing on digital image processing for the front viewcamera images collected implementations include color and texture feature extraction,feature extraction methods and different types of characterization capabilities on theground for analysis.Visual signal feature extraction based on the K-means method adaptive clusteringalgorithm to generate dynamic training samples to machine learning, morphologicalprocessing and online learning methods to achieve the target and non-target terrestrialsurface area region segmentation. Ground floor for the selected target area, theestablishment of the visual features and the mapping between the ground type.Established based on the type of ground support vector machine classification real timeidentification model and statistical model based on sample data matching algorithmsclassify the type of identification on the ground, establish a complete adaptive real-timeonline classified ground type identification method.Combined with the visual signal and vibration signal type classificationrecognition on the ground advantages and disadvantages, analyze the feasibility offusion of the two signals, the signal recognition result of vibration signal characteristicsof visual training, visual signal characteristics and the vibration signal feature fusion,and visual signal recognition results with vibration signal recognition result fusionmethod and analyze the effectiveness of several fusion method is established based onthe integration of visual and ground vibration signal type identification method.In tracked mobile robot, based on the type of ground to build real-timeclassification and experimental study identifies a platform to realize visual signal andvibration signal feature extraction and processing, access to a variety of ground types ofclassification accuracy and processing speed and other indicators have been establishedonline identification model for performance analysis. Conduct experimental analys is across multiple types of ground identification model robustness and accuracy andsensitivity as well as the balance between fast and accurately by measuring the rate offusion methods for achieving convergence assessment and analysis. Implementcomprehensive experiments on the ground type classification identification systemstability testing.