Visual Odometry Using Tiny Features for AGV Localization in Factory Environment

Navigation of automatic guided vehicles for transporting components inside a factory needs accurate estimates about vehicle’s position and orientation. The global positioning system fails to offer vehicle’s position in-door due to the structure of the building blocking the electromagnetic wave from the satellites. Electronic compass and radio positioning equipment are interfered severely by electromagnetic signals generated from electrical machines operating in the factory. This thesis presents a visual odometry which extracts tiny image features of the floor to estimate the displacement and orientation of the vehicle. The video camera which has an illuminating source points toward the floor to avoid disturbances resulted from reflection or shadow. We employ the scale-invariant feature transform algorithm to extract tiny image features from the image captured. Feature matching between two consecutive images enables using the triangulation method to calculate the displacement of the vehicle. Integrating the displacement data and using the rotating center of the vehicle lead to estimates of the position and orientation of the vehicle. The accuracy of estimates satisfies the requirements of AGV for factory use, and the result does not influenced by wheel slip or rough floor. Experiments under several scenarios show that path plan and modification can be achieved on the central control station. Position error is less than twenty centimeters for a path shorter than eight meters. For constant operations, the visual odometry needs a position-calibration system to eliminate accumulated errors.