Efficient Model-Based Image Understanding Method for an Autonomous Vehicle

摘要

The problem discussed in this dissertation is the development of an efficientmethod for visual navigation of autonomous vehicles. The approach is to significantly reduce the expensive computational time of landmark detection by straight-edge features. A novel, fast straight-edge-detection method for use in autonomous vehicle navigation and other image-understanding applications is presented. Straight edges in gray-scale images are detected using a new direction-control1ed edge tracking method, which gives precise estimate of the endpoints. To significantly reduce the number of exhaustive pixel computations, a random-hitting method using a pseudo-random number generator is proposed. Only if a generated pixel is significant do we start tracking the edge containing that pixel. To overcome the noisy gradient direction information, a robust least-squares linear fitting method is used to control the tracking process. The results of the algorithm show how it is efficient for landmark detection, which is important for motion control of autonomous vehicles. Thus the new method is implemented as a component of the image-understanding system in the autonomous mobile robot Yamabico-11 at the Naval Postgraduate School. An efficient world-modeling method based on the 2D model of the environment of the vehicle, including the heights of vertical edges in the environment, is presented. This modeling method is implemented with the new edge-detection method to improve the efficiency of the pose-determination algorithm (pose is a combination of the position and orientation of the camera), which is an essential task in the area of autonomous vehicle navigation.