Object Perception For Intelligent Service Robot

With the development of society and population ageing phenomenon becoming more severe, and the continuous improvement of the human cost, the demand for intelligent service robots is increasingly urgent. As one of the basic functions of intelligent service robots, object recognition has been a hot research topic in the field of robot vision. However, because of the complexity of the problem itself and the uncertainty of the environment, to build a reliable real-time object recognition system for mobile robots is still a challenging task.The starting point and ultimate goal of this research is hoping to realize such a system, which provides basic functionality for service robots to manipulate objects and a higher level of perception decisions, in order to meet the needs of the majority of practical applications. To achieve this goal, this thesis conducts more thorough research to the relevant issues of object recognition system, the major research contents and innovations are as follows:(1) An algorithm for automatic object segmentation is proposed. The difference between service robot and traditional industrial robot is that the former requires intelligence and autonomy. The vast majority of existing robot object recognition systems require a lot of human involvement in the process of learning or training objects, where most of the work carries on the manual object segmentation. In this thesis, through ingenious combination of the3-dimensional plane segmentation and GrabCut algorithm, an automatic segmentation tool for plane objects is implemented, which can get comparatively accurate segmentation results. Coupled with the autonomous mobility, robots can realize automatic modeling of objects.(2) A construction method of object recognition system effectively combine multiple features is proposed. Simple combinations of features not only fail to improve the accuracy of object recognition, but also reduce the overall system speed. At the beginning of the proposed approach in system design fully consider trade-off between processing cost and recognition accuracy. The goal is both to achieve full use of clues provided by a variety of features to improve the recognition rate, and to meet real-time requirements. Under the guidance of this method, this thesis builds a modular object recognition system for service robot, which combines the LINE-MOD multimodal template detection, color histogram comparison and local feature matching three kinds of object recognition technology. The experiment proved that the object recognition system has a high recall and recognition accuracy. With flexible architecture, the system allows quickly and easily replaced or increased a recognition module to further improve system performance. In order to solve the problem of threshold setting difficulty in modular recognition system, this thesis proposes a dynamic threshold adjustment algorithm. As long as the adjustment interval of each threshold is presented, the algorithm automatically determines a set of appropriate values for thresholds during the recognition process. Experimental results show this algorithm can improve the accuracy of recognition system and maintain a high recall rate in the meantime.(3) From the configuration of visual sensors, the realization of the automatic object segmentation tool and the automatic object modeling, and then to the combination design of object recognition modules and the parameter optimization design of recognition system, this thesis builds a complete object recognition system for intelligent service robot. The system has successfully completed object recognition tasks in dozens of actual scenes, providing supports for service robots to fulfil more advanced and more complex tasks, which fully demonstrated the stability and reliability of the system.