Learning to Use Scene Context for Object Classification in Surveillance

Object classification in far-field outdoor video surveillance is a challenging problem because of low resolution, presence of shadows and projective image distortion. We present a classification system that learns to use scene context variables (such as position and orientation of objects), in addition to object variables (such as shape and size), to improve its performance in an arbitrary scene. The key feature of our system is that it needs only a small number of labeled examples from a few scenes, but extends well to novel scenes by extracting contextual information from unlabeled data. A three-step bootstrapping algorithm for adapting classifiers to a novel scene is proposed. A baseline classifier is first trained using object-specific features. This classifier is applied to a novel scene to calculate ‘labels’ for unlabeled data. Some of these labels are then incorporated into the training set of a scene-dependent classifier which uses both object and context features. Experimental results demonstrate the effectiveness of our adaptive classifiers for multiple urban scenes.