Problem description:
Perception, such as recognition of the environment and objects, is crucially important in robotics. Especially in tasks like grasping and navigation, stable retrieval of object’s identity, location and other properties in real 3D scene is needed.
So far, different kinds of 3D depth features have been widely used to reconstruct the object, recognize and retrieve the object pose in the real 3D point clouds [1]. These features are only based on the 3D information of objects, whenas 2D imagery also plays an important role of object recognition. For example, it is not easy to get the full point clouds of shiny and transparent objects, or to distinguish items with same shape information such as a light Coke bottle and a Cola bottle.
With the development of new sensor technology, well combined colored 3D point clouds will be given. These data provide the colored shape and texture information of objects in a real environment. The color combined 3D geometric features of desired object models will enhance not only the accuracy but also the speed of recognition process [2].
As the goal of this thesis, the model-based recognition of objects in a dynamics environment should be achieved. The colored 3D features database of objects should be built based on existing textured 3D object models. In recognition part, one feature vector should be extracted from the a query part in real scene. After the computation of the similarity score between the database of objects and real query, the identity and the pose of object will be retrieved.
Tasks:
- Analytical comparison of different kinds of 3d features;
- Colored geometrical features database building;
- Development and implementation of an algorithm for model-based recognition in real scene;
- Optimization for real-time application and experimental evaluation.
Bibliography:
[1] Rusu,Radu Bogdan, Bradski Gary, Thibaux Romain, and Hsu John, ”Fast 3D Recognition and Pose Using the Viewpoint Feature Histogram”, Proceedings of the 23rd IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 10/2010,Taipei, Taiwan.
[2] Asako Kanezaki, Hideki Nakayama, Tatsuya Harada, and Yasuo Kuniyoshi, ”High-speed 3D Object Recognition Using Additive Features in A linear Subspace”, 2010 IEEE International Conference on Robotics and Automation (ICRA2010), pp.3128–3134, 2010.
Prerequisites:
- Good programming skills in C++
Helpful but not required:
Contact: