Problem statement

In this thesis we are proposing a method for classification of vehicles into car makes such as Škoda 120, Škoda Favorit, etc. As the input we get an image which contains the car sample from frontal view alone or with other background scenery and other objects. In the second case we expect that a number-plate was correctly detected. As an output we classify the input image sample as a certain car make. We do not expect a high number of classes since there is a limited number of car makes which can observed on Czech roads. We expect the response within seconds, not longer (minutes, hours, ...). Since we expect relatively short response time we chose fast methods for each component of our system. We do not expect occlusion in images expect light occlusion like snowing, etc.

The recognition system proposed in this thesis is based on using a robust SIFT descriptor [25] and its variations that we introduced for feature extraction . Recognition process starts with the car localization in the image. We assume the car number-plate was detected or that the image contains the car image only. In the first case we use the scale and location of the number-plate to define Region of Interest in the image from which the features are extracted. Feature vectors are finally classified using nearest neighbour algorithm and Fisher's linear discriminant. Different system configurations are tested on two databases: database containing car images, database containing truck images.