A Novel Feature Selection in Vehicle Detection Through the Selection of Dominant Patterns of Histograms of Oriented Gradients (DPHOG)

Abstract

This paper proposes a novel method that addresses the selection of the dominant patterns of the histograms of oriented gradients (DPHOGs) in vehicle detection. HOG features lead to an expensive classification with high misclassification rates since HOG generates a long vector containing both redundant and ambiguous features (similarities between the vehicle and non-vehicle images). Several modifications of HOG were proposed to resolve these issues such as the vertical histograms of oriented gradient and one that includes position and intensity with HOG; however, these methods still contain some ambiguous features. A feature selection method can exclude these ambiguous features, allowing for better classification rates and a reduction in classification times. The proposed method uses the ideal vectors of the vehicle and non-vehicles images for selecting features in dominant patterns. The segments indicating the differences between the vehicle and non-vehicle classes are the dominant patterns, in which the length of the feature vector is shortened. We performed DPHOG on three standard datasets, in which the kernel extreme learning machine, the support vector machine, K-nearest neighbor, random forest, and deep neural network were used as classifiers. We then compared the performance of the DPHOG with eight well-known feature selection methods and three existing feature extraction methods for vehicle detection. In evaluations with each comparative method concerning the accuracy, true positive, false positive, and F1-score, the DPHOG presented the highest performances with less running time in each dataset.