The visual terrain classification algorithm based on fast neural networks and its application

Two key issues, the extraction approach of visual terrain feature and the fast terrain classification approach, on influencing the classification accuracy have been studied firstly in order to improve the terrain classification ability of robot. In this paper, training images are convolved with the MR8 filter banks. Exemplar filter responses are chosen as texton dictionary via k-means clustering. Based on the texton dictionary, the feature histogram vectors of visual terrain images are generated by means of spatial pyramid matching method. In the algorithms of terrain classification, an Extreme Learning Machine with Tunable Activation Function (TAF-ELM) learning algorithm is proposed. The validity of this approach has been verified by applying the visual terrain classification approach is applied to the feature classification of terrain images based on combination of the feature extraction method of terrain images with the fast TAF-ELM learning algorithm. And thus the simulation results show that the remarkable improvement of the approach can improve the classification rate accuracy of terrain images with good efficiency.