Skip to Main Content
In multi-camera surveillance systems, it is important to track the same person across multiple cameras. It is also desirable to recognize the individuals who have been previously observed in a single-camera system. The method that represents a object image using a bag of visual words has been commonly used in image retrieval applications. For recognizing people, it can outperform the methods mainly based on global appearance like color histogram, and fit better to low-quality images compared to biometric features such as face and gait. In this paper we study the details in feature extraction, vocabulary building and classifier learning of the bag-of-features approach for classifying tracks of different individuals. Based on this approach, we design a online system applying incremental support vector machine learning with a decision scheme to distinguish reoccurrences from new targets. We get promising results from the evaluation with more than 100 tracks of 50 different people.
Date of Conference: 4-6 Nov. 2009