Distributed Feature Matching for Robust Object Localization in Robotic Manipulation

The feature matching algorithms are used to recognize the position of flat objects or surfaces in an image. This is particularly used for the control of autonomous robot arms for pick and place operations under monocular vision guidance systems. The problem arises where the object surface is not flat or the detected feature points belong to the different height planes. The error is much more prominent if the object is placed away from the center of the camera view that leads to projection parallax and the apparent surface geometry is distorted. The algorithm proposed in this paper identifies horizontal planes with different heights and uses feature matching on individual planes in a distributed way to find accurate position of the object. Two images of the object are required by this method to train and then find the object in a single image, this allows 3D model matching using only monocular camera without using machine learning techniques thatrequire a large dataset of training images. The algorithm works best for the multi-planar 3D objects, which have several feature pointson different height horizontal plane levels. The results have beencompared with the recent contour based feature matching method that addressed a similar problem.