Translation, Scale and Occlusion-Tolerant Recognition with Multiple Eigenspace Models

Abstract

We present a method for estimating the position and scale of occluder objects present in images of any of a number of modeled scenes. The background scenes are modeled using eigenspaces. If models of possible occluder objects are available we can then classify the detected foreground objects based on the available models. A segmentation of the object from the scene is also obtained in the process. We further handle the case when the foreground object region thus detected actually consists of mutually occluding objects and we locate, segment, de-occlude and recognize each object individually. We present numerical experiments to prove the validity of the method and demonstrate the utility of the algorithm in the detection, de-occlusion, segmentation and recognition of multiple objects in an office environment.