PixelLaser: Computing Range from Monocular Texture

The impressive advances in robotic spatial reasoning over the past decade have relied primarily on rich sensory data provided by laser range finders. Relative to cameras, however, lasers are heavy, bulky, power-hungry, and expensive. This work proposes and evaluates an image-segmentation pipeline that produces range scans from ordinary webcameras. Starting with a nearestneighbor classification of image patches, we investigate the tradeoffs in accuracy, resolution, calibration, and speed that come from estimating range-toobstacles using only single images. Experiments atop the low-cost iRobot Create platform demonstrate the accessibility and power of this pixel-based alternative to laser scans. 1 Motivation and Context Robots' spatial reasoning capabilities have matured a great deal over the past decade. Effective localization, mapping, and navigation algorithms have proven themselves in long-term autonomous driving [3,17], tour-guide [2,15], and office-navigation [8] applications. Robots’ most robust and widespread examples of spatial reasoning rely upon the ubiquitous laser range finder (LRF), which uses the time-of-flight of projected laser light to directly compute the range to nearby obstacles within the laser range finder’s field of view. Although effective, LRFs have drawbacks that have prevented them from entering the fast-growing field of low-cost commercially viable autonomous robots. As an alternative, monocular vision offers advantages relative to laser scans across several axes: cameras are less power-hungry, less heavy, less bulky, less range-limited, and, perhaps most importantly, less expensive. Less is more, however, when it comes to computation. Extracting range from pixel intensities requires far more algorithmic and computational effort than extracting range from time-of-flight. Range-from-vision approaches typically use temporal feature correspondence across a monocular image stream to deduce distance from pixels [6]. This body of work is mature, but it is worth noting that these techniques are most successful when significant spatial context is used to support feature matching. Large patches of pixels facilitate accurate and precise correspondence.