LEARNING THE OBJECT MODEL FOR AUTOMATIC DETECTION AND TRACKING FOR ROBOT GRASPING

The most natural approach for a robot to learn about a new object is the short presentation of the object either by hand or on a table. The robot should learn a model of the object and use it to later find the object again and track it. All these steps should be executed autonomously. One of our long-term goals is the usage of our system for robot grasping tasks. Following this approach we developed a method of extracting the object model using a depth image acquired through the scan of the object. The model extracted is subsequently exploited for detecting the object in the environment. After detection the approach automatically initializes a tracking method that follows the object motion to enable grasping or navigation tasks. The autonomous execution is made possible by the integration of depth and appearance data using a laser-based depth camera and a color camera. The use of both depth and colour images makes the approach robust to illumination changes and different conditions during learning the object model and then later re-detecting it. Experiments show the feasibility of the concept even in situations where the object is partially occluded and the scene is cluttered.