Towards open-ended 3D rotation and shift invariant object detection for robot companions

Robot companions need to be able to constantly acquire knowledge about new objects for instance in order to detect them in the environment. This ability is necessary since it is hard to predict what objects the robot may face in the operation phase during development. This paper presents ideas and results on two topics. The first topic is on the design of an open-ended object detection system that uses scale invariant feature key-point descriptors that are trained with a one-class radial basis function support vector machine. Unlike using other classifier-based approaches our method does not assume the number classes to be known a priori. The method is shown to be stable against full 3D rotation of the object relative to the sensor. The second issue in this paper deals with a solution on how to provide new object information on the robot. A modern range imaging sensor in conjunction with a conventional color imaging sensor is used for a first figure-background separation. The experiments presented support the basic statements in this paper. Conclusions are drawn and future work is addressed.