查看更多>>摘要：Reporters obtained the following quote from the background information supplied by the inventors: “The state of the art in machine vision or robotic vision is l argely based on cameras where the input to the sensing system is two-dimensional (2D) arrays of pixels that encode the amount of light that each pixel received over an exposure period, or on depth capture technologies (e.g., Time-of-Flight (ToF) cameras, structured light cameras, LIDAR, RADAR, or stereo cameras, to nam e a few) which provide three-dimensional (3D) point clouds, where each point in the point cloud may store its position in space with respect to the vision syste m, and may store any of a number of other data associated with the patch of refl ecting material that the point was generated from (e.g., brightness, color, rela tive radial velocity, spectral composition, to name a few). Note that 3D point c louds may be represented in “frames”, similar in spirit to the frames of images from cameras, meaning that they don’t have a fundamental representation of conti nuously evolving time. To provide useful perception output that may be used by m achine vision applications, such as, robotic planning and control systems, these 2D or 3D data often need to be processed by machine vision algorithms implement ed in software or hardware. In some cases, some machine vision systems may emplo y machine learning to determine properties or features of the world that may be salient to particular robotic tasks, such as, the location, shape orientation, m aterial properties, object classification, object motion, relative motion of the robotic system, or the like. In many cases, neither the 2D nor 3D representatio ns employed by conventional machine vision systems provide inherent/native suppo rt for continuous surface representation of objects in the environment.
NETL
NSTL