Front and Rear Vehicle Detection Using Hypothesis Generation and Verification

Abstract

Vehicle detection in traffic scenes is an important issue in driver assistance systems and self-guided vehicles that includes two stages of Hypothesis Generation (HG) and Hypothesis Verification (HV). The both stages are important and challenging. In the first stage, potential vehicles are hypothesized and in the second stage, all hypotheses are verified and classified into vehicle and non-vehicle classes. In this paper, we present a method for detecting front and rear on-road vehicles without lane information and prior knowledge about the position of the road. In the HG stage, a three-step method including shadow, texture and symmetry clues is applied. In the HV stage, we extract Pyramid Histograms of Oriented Gradients (PHOG) features from a traffic image as basic features to detect vehicles. Principle Component Analysis (PCA) is applied to these PHOG feature vectors as a dimension reduction tool to obtain the PHOG-PCA vectors. Then, we use Genetic Algorithm (GA) and linear Support Vector Machine (SVM) to improve the performance and generalization of the PHOG-PCA features. Experimental results of the proposed HV stage showed good classification accuracy of more than 97% correct classification on realistic on-road vehicle dataset images and also it has better classification accuracy in comparison with other approaches.