Stereo vision is one of the most common and robust vision algorithms used in mobile robot navigation. It has been used for mapping, localization and obstacle avoidance [8], [10], [11].
While performance of stereo has increased with growing computer power, the technique is still limited by the high computational cost of the algorithm. One reason for the lack of performance is that stereo is performed fully in each iteration of the perceptual cycle. This is wasteful because the algorithm is not taking advantage of the coherence between depth maps obtained in consecutive time intervals. We propose a method of coherent stereo as a way of increasing the speed of the stereo algorithm by using the information from the previous iterations of the algorithm and the constraints on the robot motion.
The thrust of our approach is to anticipate the changes between two consecutive stereo pairs of images. The computation is used to verify and refine the anticipated values, rather than calculate them without prior knowledge. The main computational gain is achieved by reducing the search space needed to find the correct disparity for each pixel in the image.
The changes in the stereo depth information between two time intervals can be attributed to motion of the robot and motion of objects in the environment. Our algorithm assumes a static environment, however we suggest ways of extending the method to dynamic environments.
The following is an outline of the algorithm:
At the beginning of the process stereo is computed fully because there is no prior knowledge about the scene. After the initial depth map is obtained the search space is reduced according to knowledge of the robot motion. The speedup of the algorithm depends on how accurately it is possible to determine the new depth model.
The accuracy of the new depth map depends on the error in measuring the new position of the robot. The most general constraint on robot motion is the amount it can possibly move in one time interval. If the time interval is short the change in what is observed may be very small.
The constraints on motion can be refined by knowledge of the general direction in which the robot is moving. For example, if it is known that the robot is moving forward, it should be expected that distances are getting smaller. Therefore, the search space is further reduced.
Finally, the amount of relative motion can be determined, quite accurately, by the use of odometry. In this case the uncertainty in the robot's position comes from the odometry readings, which in general are quite accurate over short distances. It should be noted that dead-reckoning suffers from accumulating errors over long periods of time [2]. Our approach, however, does not require absolute odometry readings; Rather it uses the relative change in the readings between two closely-spaced time intervals.
This paper presents the approach taken in speeding up the stereo algorithm when the constraints on motion of the robot are known. The paper compares the increases of performance achieved depending on what is known about the robot's motion.