Real-time mobile robot self-localization : a stereo vision based approach

摘要

The main focus of this thesis is vision based real time self-localization of tiny autonomous mobile robots in a known but highly dynamic environment.The problem covers tracking the position with an initial estimate to global self-localization.The localization algorithm is not dependent on the presence of artificial landmarks or special structures in the environment nor does it require that features should lie on or close to the ground plane.The algorithm enables the robot to find its initial position and to verify its location during every movement.Global position estimation in unmodified environments normally involves measuring distance to or angles between distinct features (natural landmarks) from the robot position or matching a local map constructed from sensor readings to a global map of the environment.On contrary to other localization algorithms stereo vision based depth computation is used for self-localization.A localization framework is in progress that uses trilateration based techniques whenever distinct landmark features are extracted.The trilateration based method is complemented by a sparse 3D map of the local environment constructed based on sensor data and matching it with the environment model.The stereo vision system is mounted on a pivoted head as an aid in feature exploration.Distance measurements are used as they require fewer landmarks compared to methods using angle measurements.Visual features are extracted using Gradient Based Hough Transform (GBHT), which provides the strongest groupings of collinear pixels having roughly the same edge orientation.Global self-localization is computationally slow and sometimes impossible if enough features are not available.Therefore, once the robot position is computed it is tracked with local sensors.This is fast and reasonably accurate as the accumulating error is suppressed after short intervals.Extended Kalman filter is used to fuse information from multiple heterogeneous sensors.Keeping a rough estimate of the robot position helps in features extraction and matching with the global map.Significant performance improvements have been achieved with a new hybrid method that combines the global position estimation with tracking.Simulation results for the robot environment modeling, feature extraction, depth computation, information fusion, and initial test of the framework have been reported.As such marked minimization of landmarks for vision based self-localization of robots has been achieved.