This thesis presents a visual navigation system that works robustly in general indoor environments with many dynamic obstacles, such as in supermarkets. The cost and setup time have been adequately considered for commercial deployment. Such a system is constituted by three modules, which are for global localization, obstacle map maintenance, and command generation respectively. To estimate the robot location, odometry information and visual measurements from a monocular camera are online fused via an extended Kalman filter (EKF). The monocular camera is designed to look at ceilings to avoid negative effects from dynamic obstacles. The global consistency is highlighted since a simple EKF will gradually drift due to the accumulation of estimation error, thus two complementary solutions of...[ Read more ]

This thesis presents a visual navigation system that works robustly in general indoor environments with many dynamic obstacles, such as in supermarkets. The cost and setup time have been adequately considered for commercial deployment. Such a system is constituted by three modules, which are for global localization, obstacle map maintenance, and command generation respectively. To estimate the robot location, odometry information and visual measurements from a monocular camera are online fused via an extended Kalman filter (EKF). The monocular camera is designed to look at ceilings to avoid negative effects from dynamic obstacles. The global consistency is highlighted since a simple EKF will gradually drift due to the accumulation of estimation error, thus two complementary solutions of localization based on Quick Response (QR) code markers or natural features are proposed. For the first solution, the markers do not need to be carefully installed. Although it is more robust to changes in perceptual conditions like illumination, the second solution serves as a backup plan when pre-set markers are strictly forbidden. The obstacle map is offline built and online updated according to observations from an RGB-D sensor, allowing the command generation module to figure out a globally optimal path to the target area through a classical A* algorithm, followed by a dynamic window approach (DWA) to generate local commands to execute the path. According to experiments conducted in spaces where there exist movable chairs, tables, and pedestrians, the proposed system is ready for autonomous navigation immediately after simple steps of setup, and it demonstrates real-time efficiency and robust global localization with low repeat errors, as well as the capability of obstacle avoidance.