Vision-based keyhole detection and parameter extraction in door opening task

摘要

It is important for a robot to be able to open a door in indoor environment. There are occasions when the robot has to open a locked door with a key. Problem arises when the robot tries to insert the key into the keyhole; it has to know the location and direction of the keyhole. In this paper, a procedure is developed to deal with this problem. First we detect the keyhole with the HOG features and a SVM classifier. The detection process has an average precision of 89%. Then the detected target image is processed by the opening operator in order to filter out the disturbance caused by the error in assembling. The resulting image is a binary image which contains only the pixels corresponding to the keyhole. Finally, we calculate the location of the keyhole with the image moment. The direction of the keyhole is computed with two different methods. We compared and analyzed these two method and the result shows that the calculated dominant direct has a constant error about one degree.