Automatic vehicle detection using statistical approach

Abstract

This study develops a statistical approach to the automatic detection of vehicles. Compared to traditional approaches, which consider the entire 2-dimensional vehicle region, this study uses three meaningful local features for each vehicle to perform vehicle detection. The proposed approach has a superior tolerance toward wider viewing angles and partial occlusions. Four possible models for vehicle detection are evaluated in the current training and testing processes. For the process of the best model, each local subregion projects into corresponding eigenspace and residual independent basis space with subregion position information. We further simplify the procedure steps of computing the independent component analysis (ICA) in residual space without constructing residual images in order to reduce the computational time. Then the joint probability of projection weight vectors and coefficient vectors of local subregions and positions of local subregions, is used to model the vehicle. Finally, we introduce vector quantization with a new classification method to accelerate the posterior probability calculation. © Springer-Verlag Berlin Heidelberg 2006.