Sliced curvature scale space for representing and recognizing 3D objects

Abstract

Perception plays a key role in the development of intelligent autonomous systems. In particular object recognition and registration tasks are crucial to any intelligent autonomous system such as autonomous cars or personal robots. The representation of 3D object sensor measurements largely affects the choice of higher level processing possible on the sensor data. We explore the use of scale space theory via the curvature scale space and extend it to represent 3D objects in our new SCSS (Sliced Curvature Scale Space) framework. We further develop techniques of further processing the SCSS representation including feature extraction and dimensionality reduction for use in learning frameworks. We perform an array of experiments to validate the effectiveness of our method and demonstrate recognition performance using support vector machines. The results indicate that our new representation retains the nice qualities of the original curvature scale space method while being robust and compact for 3D object representation and recognition.