Unstructured road detection in color imagery for the purpose of the automatic detection of explosive devices

Abstract

This thesis proposes a method for segmenting an unstructured dirt road in color space images using color and texture analysis. A support vector machine (SVM), or a Random Forest classifier is trained on samples of on and off road patches from a similar road. Image patches are classified at an interval of 10 pixels at a fixed horizontal distance from the vehicle. Each patch is described by the Histogram of Oriented Gradients (HOG), the Local Binary Patterns (LBP), a histogram of the three color channels, and a set of statistics are calculated on the color histograms. The classified patches are transformed to the next frame of the sequence using the scale invariant feature transform (SIFT) to reduce classification of image patches that have been classified in previous frames. Morphological opening and closing are used to transform the points into a mask, and reduce errors. Further error reduction and smoothing of the boundary is made possible by following the corners of the positive road classification with a Kalman filter. Experimental results measured against a hand segmented ground truth indicated that the algorithm can accurately segment road images given a set of training data from similar road utilizing only color imagery.