3D Object Registration and Recognition using Range Images

In recent years, retrieving semantic information from digital cameras, for instance object recognition, has become one of the hottest topics of computer vision. Since the boundaries of this problem range from recognizing objects in a range image to estimating the pose of an object from an image sequence, a variety of studies exist in the literature. Before remembering the previous approaches on the subject, it is better to refer to the definitions of the elementary attributes of the fundamental step in object recognition, feature extraction, which are repeatability under orientation, scaling, sampling and noise. Orientation invariance in computer vision is the ability, which enables a method to extract the same features from the original and the rotated (oriented) version of an image (2D or range image etc.). Since any movement in the scene or camera introduces considerable rotation in the signal, this is a basic ability which is usually satisfied in recent object recognition approaches. The critical setback for orientation invariance is the self occlusion caused by the orientation of the object with respect to the camera. Many recent studies solve this problem by partial matching algorithms, which also enable pose estimation in scenes with diverse rotation. The ability to extract features independent of their scale, namely scale invariance, is another important ability for a method. For 2D images, it is both related to object’s size and pixel resolution. Thus scale invariance in 2D also correlates with invariance under sampling, which, as it name implies, is the ability to extract similar features from similar signals with different sampling rates. However, since range images encapsulate metric information independent of the resolution they have, the notion of scale and sampling invariance is interpreted in a different manner for them. This different notion is discussed in the succeeding sections. Finally, the very basic ability of a computer vision system is its robustness to noise. Noise exists in various forms depending on the scene and sensor attributes. Since a very basic operation in image processing literature, unlimited number of approaches exits. However, methods using scale-space of signals have proven to be most robust under different types of noise.