A foveal 3D laser scanner integrating texture into range data.

摘要

Abstract. In the meantime the acquisition of dense point clouds or triangulatedsurf ace data with 3D laser range scanners is in the meantime a widely discussedtopic. Especially in the areas of w orld modelling or autonomous na vig ation reliable3D data is necessary . Mobile systems applications like collision avoidance or si-multaneous localisation and mapping (SLAM) require high data rates, dense pointclouds, constant availability of data and exact single scans. Commercial 3D laserscanner are mostly expensi ve, not mobile, too slo w or not exact enough. W e presentin this article a 3D laser range scanner which is based on a commercial 2D scannerwith a mo vable origin. In contrast to other self-b uild 3D scanners our scannerreaches high data density in heading direction and lower density at the boundaries(fo veal vision), which is important for autonomous na vig ation.K eyw ords. 3D vision, depth image, texture inte gration, w orld modelling 1. Intr oductionIn autonomous na vig ation, collision avoidance and w orld modeling are still major topicstow ards real autonom y. Open question is ho w to collect, arrange and store en vironmentinformation suitably . Depending on the task different types or amounts of data ha ve tobe collecte d. In general it can be seen that dense 3D point clouds suit well for mostapplications. Hence the task for the robotO s sensory equipment is as follo ws:¥ Scanning of a wide region of the surrounding space¥ Scanner range from some cent imeters up to several meters¥ High accurac y of the laser sca nner¥ High scanning velocity¥ Possibility to conÞgure the device for different resolutions of the point cloud¥ Easy and rob ust construction¥ F oveal vision for ha ving the hig hest resolution in the heading directionDif ferent variations of mo ving 2D scanners ha ve already been proposed. In [1], [2], [3],[4], [5] or [6] sin gle scan lines are tak en with a hori zontal mo vement. This approachcan generate a good model of the en vironm ent, but cannot avoid collisions. Depth imagebased approaches can be found in [7],[8], [9]. Here Þrst steps tow ards autonomy or semi-autonomy are tak en. The tak en depth images sho w a high resolution in (for the robot)