Extended Histogram of Gradients with Asymmetric Principal Component and Discriminant Analyses for Human Detection

Asymmetry in training sets of humans and non-humans and high dimensionality of existing features are problems plaguing human detection. As classification of humans tends to be one class versus all other classes, existing classification methods do not consider this asymmetry in the training sets which leads to sub optimal classifier performance. Furthermore, the high dimensionality of existing features hampers real-time performance of human detection and classification. In this paper, we address these 2 issues by considering Asymmetric Principal Component and Discriminant Analyses (APCDA) for use with modified Extended Histogram of Gradients (ExHoG) with Mahalanobis distance classifiers. APCDA is a dimensionality reduction and feature extraction technique which specifically addresses the asymmetrical training sets problem and improves classification performance. We modify ExHoG by performing normalization on the Histogram of Gradients block features to suppress any large peaks in gradient magnitudes. Our experimental results, using the INRIA Human data set, show that the modified ExHoG results in an improved performance over the original ExHoG. Furthermore, using APCDA and Mahalanobis distance classifiers, our method with the modified ExHoG outperforms state-of-the-art methods on the Daimler Pedestrian Classification Benchmark data set.