We describe a method to recognize moving obstacles in a wide view and in real time required when a mobile robot moves in a dynamic environment. Our method uses an omnidirectional stereo vision composed of a pair of vertically-aligned omnidirectional cameras and a PC cluster to obtain panoramic range information of 360 degrees in real time. From this range information, the robot on-line generates a free space map of the surrounding environment, and extracts objects in free space as candidates for moving obstacles. The robot makes time correspondence of candidates and estimates their position and velocity using the Kalman filter. To reduce the effect of odometry error to map generation, egomotion is estimated by comparing current and previous range data. We demonstrate the effectiveness of our method by on-line experiments.