Vision based localization of a moving target

• Main Author: Selvaraj, Prabu
• Other Authors: School of Electrical and Electronic Engineering
• Format: Theses and Dissertations
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/61627
• Institution: Nanyang Technological University

Localization is the process of making a target or robot realize its current location. It is one of the most important task in autonomous navigation of either UGV (Unmanned Ground Vehicle) or UAV (Unmanned Aerial Vehicle). Localization is the basic information for any robot navigation tasks. Without localization information, robot's path cannot be tracked. Many robotic tasks such as obstacle avoidance, path planning, indoor and outdoor exploration etc ..., relies on position information of the robot. Localization is performed by tracking robot's position throughout the robot's movement using sensors. Most researchers use Global Positioning System (GPS) as the sensor to directly obtain the robot's location. But GPS signal can be interrupted by tall buildings, dense trees etc…In those cases, we need to rely on other sensor such as Laser, IMU (Inertial Measurement Units), Camera etc…for localization. In this research work, a vision based localization technique is proposed. The significance of this vision based localization method is that it can guide the robot more reliably even in the absence of GPS signal. A Bumblebee 2 stereo camera is used in this research to localize a robot with suitable computer vision techniques. The novelty of the algorithm is that two algorithms are developed in which one is used for robust pose estimate and other is used for global localization. In the entire experiments, we use stereo camera as the only sensor. Experiments were performed using commercially available Pioneer robot as well as using few custom made robots in the lab. Suitable feature extraction and feature tracking techniques were used for obtaining the robot's current pose. Scale Invariant Feature Transform (SIFT) is used as the feature detector. These features are tracked in subsequent frames. The initial and final position of the features are calcu-lated from which rotation and translation are estimated. From the estimated rotation and translation, localization is achieved. Finally few techniques to achieve localization using single camera are discussed and methods for improving existing localization techniques are also suggested.