Vision Based Autonomous Robot Navigation: Motion Segmentation,

Abstract

The ability to acquire and respond appropriately to targets or obstacles, moving or stationary, while underway, is critical for all unmanned mobile robot applications. This is achieved by most animate systems, but has proven difficult for artificial systems. We propose that efficient and extensible solutions to the target acquisition, discrimination and maintenance problem may be found when the machine sensor-effector control algorithms emulate the mechanisms employed by biological systems. In nature, visual motion provides the basis for these functions. Because visual motion can be due either to target motion or to platform motion, a method of motion segmentation must be found. We present a solution to this problem that emulates natural strategies, and describe its implementation in an autonomous visually controlled mobile robot. The Problem of Motion Segmentation Motion segmentation is used here as the process of identifying the 3-D spatial location of a unique source of motion from the optic flows created by one or more independently moving objects in the visual field. Motion segmentation is a greater problem than identifying and localizing a source of motion because the observer himself may also be moving in a complex visual space causing all other objects, whether stationary or moving, to contribute to the optic flow though induced motion. Consider a person driving his car down a busy street. In the course of driving, the driver uses saccades to select targets upon which to fix his gaze, and pursuit eye movements to hold his gaze upon his selected target. These targets may be other moving cars, pedestrians, billboards or other objects resting on the ground. While the relative target motion is minimized by the fixation mechanisms, images of the objects in the foreground and background relative to the target will continue to contribute to the optic flow on his retina during target fixation due to the motion of the car as well as to the pursuit eye movements. This flow can be non-uniform in direction as well as speed, depending upon the distance to the objects and their position with respect to the target and the relative direction of gaze with respect to the direction of travel of the car. Still, new targets are quickly noticed in the driver’s peripheral vision, especially if they are moving independently of their surround. The mechanism that draws attention to these new targets is the focus of our interest in motion segmentation. Studies of eye movements of drivers have shown that unusual motion is a strong attractor of attention (Thomas, 1969). For example, a car that takes off from the parked position is irresistible, so is a blinking light, like a turn indicator, or a bouncing ball that appears across the traffic. While the driver most often notices objects that are themselves moving, motion in itself is inadequate to