Algorithm Research On Terrain Classification For Wheeled Robot Using Vibration Signals

Wheeled robots need to have certain autonomous recognition capabilities when traveling on the ground,and take corresponding walking solutions according to different ground types.The current mainstream research is based on visual or lidar terrain classification methods,however,these methods often do not work well when the lighting conditions are poor or when there are coverings on the ground.Therefore,this paper uses vibration signals as input features to realize the terrain classification of wheeled robots,the vibration signal generated by direct contact between the wheel and the ground can truly reflect the information of the ground bearing layer,it is an important supplement to the terrain classification research of wheeled robots.The research on terrain classification methods in this paper mainly includes three parts,one is to give a feature extraction method,the second is to design a classifier method based on CNN and KELM,the third is to give the feature-level fusion algorithm and a decision-level fusion algorithm based on D-S evidence theory of two input features.This paper introduces the data acquisition experiment of the original vibration signal of the wheeled robot.Let the wheeled robot travel at a constant speed of 5 speeds on 5 experimental grounds of sand,gravel,grass,soil and asphalt,the vibration signals generated by the wheeled robot traveling on the experimental ground can be collected by installing a three-way accelerometer and a z-direction microphone on the wheel arm,it can provide experimental data for subsequent classification studies.The vibration signal is preprocessed according to the actual research environment,and the short-time Fourier transform(STFT)is used to extract the feature of the experimental data to obtain the time-frequency map feature.The image texture features of the time-frequency map are further extracted to obtain the local binary pattern(LBP)feature map and Gabor wavelet feature map of the time-frequency map of the experimental data for the subsequent classification algorithm research.This paper introduces the model structure of the VGG-16 convolutional neural network(CNN)and simplifies the structure according to the actual sample environment of this study,the simplified CNN model is used to classify the five experimental grounds at five speeds based on LBP feature map and Gabor wavelet feature map;In order to solve the problem of insufficient classification performance of the fully connected layer in the CNN structure,extract the feature vector of the hidden layer in the middle of the fully connected layer as a new input,and use the kernel extreme learning machine(KELM)algorithm for ground classification.Set up a control group,use extreme learning machine(ELM)and support vector machine(SVM)for classification,compare and analyze their performance.Considering the difference of the vibration information extracted by the two image texture features,in order to improve the classification accuracy,This paper presents the feature-level fusion method and decision-level fusion method based on LBP feature map and Gabor wavelet feature map.Among them,feature-level fusion is to stitch the two features when extracting CNN fully connected layer features,decision-level fusion is to use D-S evidence theory to make decision fusion of the output probabilities of the two features after classification,and comparative analysis of classification accuracy and real-time performance of the above two fusion algorithms.