Principal component analysis-based learning for preceding vehicle classification

This study presents a new scheme for cluster generation and classification of preceding vehicles from images. The proposed clustering algorithm models the distribution of vehicle images using `vehicle´ clusters. `Non-vehicle´ clusters are generated by modelling the distribution of non-vehicle images. The clusters are created using K-means clustering algorithm. Hierarchically related nested eigenspaces are acquired to reassign the patterns of each cluster. An appropriate classifier is obtained to classify the vehicles based on the `distance-from-feature-space´ measurement. The eigenspaces of vehicle clusters together with non-vehicle clusters are used for classification. The approach of modelling the distribution of vehicle and non-vehicle images and the choice of the classifier used are investigated through experiments thoroughly. Comparison on the performance of the proposed scheme is made with that of MultiClustered Modified Quadratic Discriminant Function approach of categorising the preceding vehicles. The superior performance of the proposed scheme is clearly illustrated through the classification results.