Object detection and recognition in images, is still subject to extensive researches because of the difficulty of the task: changes in light or camera displacements may lead to images in which a same 3D object can appear in multiple different 2D aspects.
Hence, we plan to develop some original image processing techniques in order to detect the similarities between the different views of a same object or a same class of objets.
An example of application is mobile robot navigation : a robot, with an embedded camera, acquires different 2D images by moving in its environment, and then concludes in real time that it is placed in a room with one table, four chairs, two plants, etc.
We have firstly focused on Zernike moments to recognize specific 3D objects in a 2D scene: robustness and generalization ability are excellent, but the method is quite slow and no so robust to occlusion. We have then studied a fast random tree keypoints-based method, however this approach can only work with textured objects. Hence, we are working on a more generical approach, which would allow to modelize a class of 3D objects using at once contours, keypoints and texture. We want to inspire from existing hierarchical and bio-inspired methods, both especially fast and efficient.