At the recent International Robots & Vision Show, Evolution Robotics (Pasadena, CA, USA; www.evolution.com) reported that it has licensed its Evolution Robotics Software Platform (ERSP) to Bandai (Tokyo, Japan; www.bandai.co.jp) for a personal robot modeled on Doraemon, a cat-like robot. This robot will be developed by 2005 and will be targeted as an entertainment and educational robot in Japan and Asia. The software architecture behind the ERSP is an open platform that enables integration of software, hardware, and complex applications for personal robots. Software modules include autonomous navigation, visual object recognition, and human-robot interaction using speech and gestures. To broaden the appeal of its products to industrial users, Evolution Robotics demonstrated its ERSP suite of tools to OEMs and systems developers at the show.
"Of the four technology modules for vision, navigation, interaction, and architecture that are supplied with the ERSP, most of the company's technology development is focused in the area of computer vision," says Jennifer McNally, senior director of marketing at Evolution Robotics. Vision technology proves especially useful for navigation, manipulation, and human-robot interaction.
"The greatest challenge for vision technology is the development of algorithms that reliably and efficiently solve problems in realistic settings that require limited computing power," McNally adds. "The main strengths of the object-recognition module lie in its robustness in providing recognition in environments where lighting can change dramatically.
"Our ERSP vision software can recognize an object even if it is at a different orientation with respect to the camera." As well, objects can be recognized at different distances from the camera. Like other machine-vision software from companies such as Matrox Imaging (Dorval, Quebec, Canada; www.matrox.com/imaging) and Cognex Corp. (Natick, MA, USA; www.cognex.com), Evolution Robotics' software also handles occlusions of up to 90%, depending on the object and distance from the camera.
To train the object-recognition system, users can capture one or more images of the object of interest and load them into the software's database. The software then analyzes the object's image and finds thousands of unique and local features to build an internal model of the object. The ERSP recognition speed is a logarithmic function of the number of objects in the database.
"Hence," says McNally, "the object library can scale to hundreds or even thousands of objects without a significant increase in computational requirements." The recognition frame rate is proportional to CPU power and image resolution.
As a benchmark, using images of 320 × 240-pixel resolution, the recognition algorithm runs at 5 frames/s on an 850-MHz Pentium III and at 3 frames/s at 80 × 66 pixels on a 100-MHz, 32-bit processor. Training is accomplished by capturing one or more images of an object from several points of view. For planar objects only the front and rear views are necessary. For 3-D objects several views covering all facets of the object are necessary. A small subset of the 1000 captured features identifies the object and provides its name and x, y, z, and theta locations.
