Research On Visual Tracking Technology Of Mobile Robot With Binocular Vision System

The thesis will focus on the design and implementation of motion tracking system under the experimental platform of mobile robot MT-AR, which is installed with binocular vision system. The goal of our study is to make robot sensitive and responsive to tracking object in normal tracking condition. The thesis will describe the inner tracking method of mobile robot and dive into the design of tracking and controlling system of the binocular vision robot.Different from the monocular vision robot system, the robot used in the thesis belongs to the mobile robot platform equipped with the binocular vision system. The inner tracking method uses the semi-supervised model, the prior overlap information of sampled example is exploited to improve the multiple instance learning(MIL) method. The bag model of discriminative function in MIL is directly constructed on instance level using Fisher linear discriminative. To enhance the tracking robustness of mobile robot, one retrieval mechanism is employed in the thesis, the object could be retrieved even when it gets lost in the scene.In traditional “tracking-by-detection” tracking, the self-learning tracking procedure of tracking system often brings about the bad result of the tracking drift. To solve the problem, an improved multiple instance learning tracking based on semi-supervised learning tracking(MILFLD) is employed in the thesis. This model blends both the labeled and unlabeled data leveraging the merits of overlap information of prior sample knowledge, which could get rid of tracking drift about traditional multiple instance learning implementation. Upon its inner components, the discriminative model of bag model utilizes the Fisher linear discriminative function, the weak classifier is constructed on the instance level. Last but not least, the selection of weak classifier is in the perspective of gradient decrease of error propagation at the cost of the maximum decrease of loss function, thus the discriminative capacity of strong classifier would still be discriminative in the future frame. To testify the tracking accuracy and robustness of inner method used in robot, numerous experiments have been performed under diverse environment with MILFLD method. The experimental results show that our method can handle motion tracking in various scenes, it is very stable and robust no matter under the challenging scenarios of scale and illumination variations, fast motion, occlusions, partial occlusions and etc. To demonstrate the tracking accuracy of mobile robot, numerous experiments of robot tracking have also been performed in the real world. The experimental results show that, the robot achieves the goal of having a real-time automatic tracking and is very responsive for the motion response, the mobile robot can handle the scenes of fast motion, occlusions, object disappearance and illumination variations very well.