The Combination of Different Cell Sizes of HOG with KELM for Vehicle Detection

Abstract

HOG has been developed successfully in many intelligent vehicle detection systems. HOG still has interesting problems that consist of (i) redundant features and (ii) ambiguous features (similarities between vehicles and non-vehicles), which problems have an effect on time computation and misclassification. The vertical direction of HOG method (V-HOG) and adding the position of orientation bins and intensity features (πHOG) improve the problems of HOG; but they produce redundant and ambiguous features in various regions of vehicles. This paper proposes a new method for improving the performance of HOG that has flexibility in various regions of vehicles. The proposed method used combines different sized cells of HOG that is called CDC-HOG. The CDC-HOG were conducted on a GTI dataset, which consists of four regions (far, front, left, and right regions). The CDC-HOG is compared with HOG, V-HOG, πHOG, and PHOG; uses the kernel extreme learning machine (KELM), and supports vector machine (SVM) for evaluating features. The CDC-HOG with KELM produced the highest performance in terms of accuracy, true positive rate, and false positive rate for all regions.