Object Recognition Based On Vision For Mobile Robot

Using visual sensors for target recognition research is a hot issue in robotics, the key question is how to extract the features effectively from the video image with a huge amount of information, as to say, how to find a high speed and effective object recognition algorithms to complete the object recognition task in a complex natural environment in real time and meet the mobility of the moving robots.This dissertation studies the mobile robot target recognition problem based on monocular vision. On the basis of analyzing large numbers of target recognition algorithms, the dissertation first studies features extraction, special color object recognition based on improved color image segmentation algorithm and object localization based on ground constraint. Then ORB local feature extraction algorithm for object recognition is studied. An improved template matching method with dominant gradient orientation is studied for texture-Less objects recognition problems. With this method, an object can be recognized from multi-angle in real-time and different objects can be recognized in the same time. The major work is as follows:Firstly, the dissertation introduces the background and significance about monocular-based object recognition for mobile robot and reviews the state-of-art of vision-based object recognition and the main problems. The main work and framework of this dissertation is given.Secondly, in the initial stages of research, we propose to use images global characteristics (color), and with an improved color image segmentation algorithm based on HSV model, realize special color object recognition. Further the target can be localized with the ground constrained method. This method is simple and suitable for the situations such as the precision requirement is low, the background is relatively simple and the object’s color is prominent.Thirdly, aiming at a large computational burden, high dimension feature vector and complex matching process imposed by SIFT or SURF feature extract algorithm, ORB (Oriented FAST and Rotated BRIEF) is studied. The algorithm is built on the well-known FAST key point detector and the recently-developed BRIEF descriptor. The comparison experiments have been done to test the properties of ORB relative to SIFT and SURF. The experiment results validate that the speed of ORB can reach5～10times than the SURF algorithm and the algorithm can control the number of extraction of feature points to meet the real-time and efficient identification requirements, so it provides a new method for mobile robot target recognition problem.Fourthly, a real-time template matching recognition algorithm is studied for object recognition algorithm compared to based on local feature extraction algorithm which is difficult for texture-less object recognition and the object can not been recognized from different perspective. The template representation is designed to be robust to small image transformations and it can create different templates from one template by Affine Invariant Transform. This robustness based on dominant gradient orientations allows to test only a small subset of all possible pixel locations when parsing the image, and to represent an object with a limited set of templates. A binary representation makes evaluation very fast and a branch-and-bound approach efficiently scans the image, it can detect texture-less objects in complex situations and recognize the same object from different perspective.Finally, conclusions are given with recommendation for future work.