Optimized template trees for appearance based object recognition

Appearance-based 3D object recognition is becoming popular for the recognition of faces and unstructured objects in new applications, such as mobile service robots in health care environments. Typical recognition problems are a cluttered background, the need of real-time recognition for in-the-loop applications, like object recognition and grasping by a robot, and a large search area, often the full camera image. In our approach a search window is shifted across the image in a way similar to cross-correlation and after each shift the content of the search window is classified by use of a template tree. To achieve real-time capability, adaptation of template resolution and step-size between consecutive search windows are integrated within the tree structure. The robustness of the tree classifier is significantly increased by optimization of the classifiers within the tree nodes by use of information theory methods