Moving Vehicle Classification Using Eigenspace

A novel moving vehicle classification technique is introduced in this paper. The proposed method considers only particular viewpoint(s) of a car for classifying similar car and/or non-car. We obtain three different views (rear, side and front views) of cars and develop an eigendatabase. This paper shows that eigenspaces are well separated with respect to the car views and, therefore, car and non-car can easily be classified using the respective eigen subspaces. We then generalize the respective eigenspaces taking a mean of the respective feature points. Since the car eigenspace finally contains only the feature of obtained views, it makes our approach very simple and computationally inexpensive. For the classification, we also present a new matching algorithm based on eigendimension for classifying car and non-car. The proposed classifier is compared with Euclidean and Mahalanobis distance-based classifier, and experimental results from the real world scenes show the effectiveness of the proposed method