Sep. 26 2008 10:00AM

Sep. 26 2008 11:00AM

The overview talk about various research interests of our guest Elizabeth González. Elizabeth González is a Ph.D student from Castilla La Mancha university, Spain. She is working in 3D object recognition applied to grasping tasks in robotic applications and the 3D objects are represented by their appearance (appearance-based).

Firstly, a comparative study between different 2D shape representation methods (Fourier Descriptors, Hu moments, Invariant Integral, Shape Contex, Zernike moments and Flusser moments) and identification methods (deterministic and stochastic similarity measures) applied to 3D object recognition, is discussed. In the experimental tests the recognition rate and the uncertainty and ambiguity in cases of geometric transformations, viewpoints changes and noise are evaluated.

The second part of the talk is focused on improvement of the 3D object recognition systems using the active recognition paradigm. Two different active recognition strategies will be presented. They use a spherical structure for the representation of 3D object views. One of them presents a 3D object recognition/pose strategy based on Fourier descriptors clustering for silhouettes. The method consists of two parts. Firstly, an off-line process calculates and stores a clustered Fourier descriptors database corresponding to the silhouettes of the synthetic model of the object viewed from multiple viewpoints. Next, an on-line process solves the recognition/pose problem for an object that is sensed by a camera placed at the end of a robotic arm. The method solves the ambiguity problem - due to object symmetries or similar projections belonging to different objects - by taking minimum number of additional views of the scene which are selected through a heuristic next best view (NBV) algorithm. The second recognition system describes a new strategy of optimal viewpoint selection for 3D object recognition purposes. Feature vectors of an object viewed from specific viewpoints are mapped on the nodes of a tessellated sphere, which we call D-Sphere. The next best view (NBV) is established on the D-Sphere and is defined as the view that yields the maximum dissimilarity among the set of candidate objects (hypotheses). The purpose of finding a NBV is discriminating quickly between hypotheses. Our method aligns all hypotheses into a tessellated sphere. Then it calculates the viewpoint (node), which exhibits the maximum dissimilarity among the corresponding views of the candidate objects. The node values of the D-Sphere are weighted taking into account the kinematics costs involved in the movement of the robot arm that carries the camera.