Research On Terrain Classification For Robots Based On Restricted Boltzmann Machine

When the mobile robot moves in a complex environment,its autonomy largely determines whether the task is carried out successfully or not.The extremely important factor that determines the autonomous mobility of mobile robots is the ability of terrain detection and recognition.Mobile robots must have the ability to identify and get through safely unknown terrain,so the terrain classification have been an important direction for machine learning.At present,most of the researches on the terrain identification and classification are based on radar and vision,but this method is easily influenced by illumination and surface covering,and is difficult to identify soft terrain.So this paper adopts the method based on vibration signal.Different from non-contact method,the vibration signal can reflect the bearing layer information of terrains,and is an important supplement to the visual terrain classification method.The research work will be carried out from three aspects: the first is to present the feature extraction methods;the second is the design of classifier based on classification restricted boltzmann machine;the last is fusion algorithm based on D-S evidence theory.This paper introduces the data acquisition experiment system and experiment process of the data acquisition experiment.A mobile robot that installed a three-direction accelerometer and a microphone in z direction on the four wheel arms respectively is used to acquisite the vibration signals of wheel-terrain interaction to classify the terrains by traversing on sand,gravel,grass,soil and asphalt terrains with five different velocities.The experiment provided data support for the latter research of classification algorithm in chapter three to five.The original experimental data were divided into short data segment.The Double Hidden Layer Restricted Boltzmann Machine(DHL-RBM)network model was constructed and it was successfully used to extract the feature of the vibration signal;Using Singular Value Decomposition(SVD)and Time Amplitude Domain Analisys(TADA)feature extraction method,feature of original vibration signal was extracted to obtain a feature sample set,which was used to study the post order terrain classification algorithm.This paper introduces the model of the Classification Restricted Boltzmann Machine(CRBM).And taking into account that the feature of DHL-RBM extraction is the probabilisticform,the feature of SVD and TADA extraction is the real value form,the standard CRBM and Gaussian CRBM ground classification algorithm have been used at five speeds based on DHL-RBM,SVD and TADA of the three feature extraction methods for five kinds of terrains,which probabilistic output form of terrain classification results can be obtained.The real-time analysis of classification methods was carried out.Considering the three features can reflect the different aspects of the data itself and the differences in classification results of classification restricted boltzmann machine on the 3feature extraction methods,the classification method of fusing two and three single H-RBMs respectively based on D-S evidence theory was proposed to improve the accuracy.The real-time analysis of fusion methods was carried out.The proposed algorithms have been validated by corresponding experiments.