Research On Terrain Classification For Mobile Robot Based On Wavelet And SVM-kNN

The Mobile Robot usually serves as an important platform for the human to explore the unknown environment. So the terrain classification of mobile robot becomes a hotspot of research. The current classification is mainly based on ground radar and laser vision, which have poor ability to identify the terrain with the covering, especially the vision-based methods require a higher light. The methonds based on vibration can read the characteristic of terrain when there is no enough light. It is very important for improving the autonomous performance of mobile robot. So the paper will research on terrain feature extraction and classifier designing.Firstly, Building the experimental data collection platform.when a mobile robot with six speeds travels across five kinds of terrains, extract the raw vibration signal from robot-terrain interaction. It will be ready for feature extraction and classification.Secondly, the paper adopts wavelet packet transform to processed vibration signal,because wavelet packet transform is better at processing the local signal in detail. F wavelet packet PCA method and the logarithm of wavelet packet energy entropy method are proposed to extract terrain features. Wavelet packet PCA method combines the feature evaluation and principal component analysis for the amplitude characteristics.Finally, when more than two kinds of terrains have the same number of maximum votes,one-against-one SVM can't give the classification result; For Directed Acyclic Graph SVM(DAGSVM), the errors accumulated and passed down, however, the problem can be solved via effective classification path planning. The paper introduces the algorithms that combine k -nearest neighbors (kNN) algorithm and SVM to improve performance of SVM. The paper separately uses two feature extraction menthods based on wavelet packet, kNN, the combination of kNN and SVM to classify the terrains, and gives out comparison and anaslysis of the classification result.Based on the classification results of measured data, the prosposed method have been verified effective.