Pedestrian classification using principal eigenspaces

The proposed work describes a distribution-based approach for recognizing people in images. The methodology involves pattern classification using first and second order statistics in Principal Component Analysis (PCA)-based clustering framework. Unknown distributions of pedestrian and non-pedestrian patterns are approximated by learning the first and second order statistics of the sample images. Normalized Mahalanobis distance measure is used as closeness measure for clustering and as discriminant measure for classification. Experimental results on real images are given to demonstrate the performance of the proposed method. The classification results are found to be as good when compared with Modified Quadratic Discriminant Function (MQDF)-based clustering and classification.