Dangerous object recognition for visual surveillance

In this paper, we address a critical task, i.e. dangerous object recognition for web surveillance system. Instead of investigating how to define an object is dangerous by analyzing human's activities (e.g. leaving a bomb in public places), our research focus on how to capture the dangerous object immediately when he/she appears under surveillance camera again, to stop him/her do more bad things. Different with the existing template and feature matching based methods, we solve the dangerous object recognition problem by a classification based method. We train a SVM classifier by learning "bag of words" based dangerous and non-dangerous object representation. For obtaining more discriminative object descriptors, we fuse color and texture, two low level image features, to generate descriptors under "bag of words" frame. We evaluate the proposed method, along with template and feature matching methods. The experimental results validate our method.