Traversable Region Detection Based on One-class SVM and Multi-visual Features Fusion

For the difficulty in obtaining the complete non-traversable region samples,a traversable region detection method based on one-class SVM(support vector machine) is proposed to improve the adaptability of algorithms in different scenes.This article formulates traversability detection as a one-class classification problem for the first time.An improved feature extraction method is proposed with the fusion of color and texture.Image data of every color channels are transformed by discrete cosine transform(DCT),then the DCT coefficients are decomposed using pyramid decomposition.Mean and variance in each decomposition are used to describe characteristic window.Traversable region pattern is generated by training the traversable samples using one-class SVM.Experiments show that the algorithm recognizes new data well,and performs with high detection accuracy and low abused detection rate.