3D Object Recognition and Facial Identification Using Time-averaged Single-views from Time-of-flight 3D Depth-Camera
Résumé
We report here on feasibility evaluation experiments for 3D object recognition and person facial identification from single-view on real depth images acquired with an “off-the-shelf” 3D time-of-flight depth camera. Our methodology is the following: for each person or object, we perform 2 independent recordings, one used for learning and the other one for test purposes. For each recorded frame, a 3D-mesh is computed by simple triangulation from the filtered depth image. The feature we use for recognition is the normalized histogram of directions of normal vectors to the 3D-mesh facets. We consider each training frame as a separate example, and the training is done with a multilayer perceptron with 1 hidden layer. For our 3D person facial identification experiments, 3 different persons were used, and we obtain a global correct rank-1 recognition rate of up to 80%, measured on test frames from an independent 3D video. For our 3D object recognition experiment, we have considered 3 different objects, and obtain a correct single-frame recognition rate of 95%, and checked that the method is quite robust to variation of distance from depth camera to object. These first experiments show that 3D object recognition or 3D face identification, with a time-of-flight 3D camera, seems feasible, despite the high level of noise in the obtained real depth images.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...