Context-based object recognition for door detection

This paper proposes a new method to detect doors using context-based object recognition. Particularly, in order to improve the efficiency of object recognition, we utilize robotic context such as the robot's viewpoint and the average height of doorknobs. The robotic context is used to make a region of interest in a captured image which reduces both the computational time and false-postive rate in the object recognition process. In addition, we employ shape features for object recognition which makes our method more robust to appearance changes than others using texture features like SIFTs and SURFs. We implemented a door detection system on a mobile robot with a stereo camera and demonstrated in corridor environments. Here, two types of doorknobs are tested: straight (door-handle) and round (door-knob) ones. The experimental results show that our method works successfully with different kinds of doorknobs in real environments.