Efficient Template Generation for Object Classification in Video Surveillance

For object classification in video surveillance, features extracted from images are compared with a visual dictionary. The best-matching features are learned by the classifier to determine the object class. In this paper, the visual dictionary concept is extended with Interest Point Operators (IPOs). In a first experiment, the influence of using IPOs on the visual dictionary creation process is measured and optimized. Secondly, given this optimal dictionary, the computational efficiency is evaluated for the dictionary matching process. Experiments show that the creation of the dictionary is most effective when extracting features at random locations. For the dictionary matching step, the use of IPOs gives a massive improvement (factor 8) in computational efficiency, while retaining a close-to-optimal classification result.