Classifying tracked objects in far-field video surveillance

Abstract

Automated visual perception of the real world by computers requires classification of observed physical objects into semantically meaningful categories (such as 'car' or 'person'). We propose a partially-supervised learning framework for classification of moving objects-mostly vehicles and pedestrians-that are detected and tracked in a variety of far-field video sequences, captured by a static, uncalibrated camera. We introduce the use of scene-specific context features (such as image-position of objects) to improve classification performance in any given scene. At the same time, we design a scene-invariant object classifier, along with an algorithm to adapt this classifier to a new scene. Scene-specific context information is extracted through passive observation of unlabelled data. Experimental results are demonstrated in the context of outdoor visual surveillance of a wide variety of scenes.