ATTENTIONAL MANAGEMENT FOR MULTIPLE TARGET TRACKING BY A BINOCULAR VISION HEAD

摘要: The development of an attention strategy for multiple target tracking and its integration with a binocular vision head is presented. Monocular motion-based segmentation yields image regions that are further clustered into targets. emulation fovea significantly reduces the computational workload segmentation. Attentional management based on assigning interest value to each target. Its evaluation weighs disparity, number detected moving pixels in target, pixel density, velocity duration along sequence. performs saccades, ballistic motions smooth pursuit around tilt vergence axes characteristics similar those anthropomorphic visual apparatus. System performance evaluated 166 Mhz PC host unexpensive off-the- shelf hardware two situations. In first case, small targets translate independently over textured background. second, human subject undergoing rotation, non-rigid motion scale changes. A processing rate about 0.8 stereo pair per second attained. case shows stabilization capabilities translating only. reveals impact limitations system sensitivity distortions gray level patterns. It induced by interframe rotation changes illumination factor as tracked moves scene. As consequence, undesired attentional reorientations occur. However, succeeds keeping within field view such conditions though not always stabilized region fixation.