Evaluating Feature Importance for Object Classification in Visual Surveillance

Feature-based object classification, which distinguishes a moving object to human or vehicle, is important in visual surveillance. In order to improve classification performance, in addition to choosing between the classification (such as SVM, ANN etc), we have to pay attention to which subset of features to employ in the classifier. This paper describes a method to evaluate the relative importance of various features for object type classification. Starting with a given set of features, we apply the AdaBoost method and then we compute a metric which enables us to choose a good subset of the features. We apply our method to the task of distinguishing whether an image blob is a vehicle, a single human, a human group, or a bike, and we determine that shape-based feature, texture-based feature, and motion-based feature are reliable for this classification task. We validate our method by comparing with performance of ANN-based classification