The glowing halls of MERL, the Mitsubishi Electric Research Laboratory, are constantly aflame with the activity of PhDs and engineers swept up in the grand opera of scientific esoterica. But that's not all, as the facility pumps out technology, for both today and tomorrow, that has real-world implications.
As a subsidiary of Mitsubishi Electric, a $31 billion multinational company, MERL is the North American arm of the corporate R&D organization. Located in Cambridge, MA, MERL conducts application-motivated basic research in computer and communications technologies. The lab has 80 employees, most of whom hold PhDs, and some of whom maintain lecturing positions at nearby MIT.
The point of this brain trust is to generate intellectual property for other groups within the larger company. Research efforts at MERL focus on five technology areas including Computer Vision, featuring the observation of people in images; Digital Communications, featuring wired networks and wireless transmission; Digital Video, featuring encoding, decoding and analysis of video; Off the Desktop Interaction and Display, featuring novel devices and interface concepts; and Sensor and Data System, featuring novel sensors, communication and system architectures. Each of these areas is, in turn, split up into two parts, the research lab, where the blue-sky stuff happens; and the tech lab, where the scientists bring the technology wonders to other divisions so they can make products for the industry.
Within the Computer Vision section of the work fall many projects that relate to the security industry, including 3D face recognition, audiovisual event detection, camera network calibration and object detection and tracking. Jay Thornton, Computer Vision applications group manager, said that in developing these technologies, there is considerable give and take between his lab and the groups in Japan who evaluate and respond.
MERL either exchanges or delivers code to other divisions via a library that they can plug in to their prototypes. The software then gets tested and rewritten in Japan before it is put into actual products for shipment.
"We have gone through four releases of that software library," Thornton related, "because we always get feedback from the product divisions saying that it works in controlled conditions, but not always real-world conditions."
The difficulty has come in situations where the person was not directly facing the camera and where poor lighting influenced image analysis. This has taken them back to the drawing board.
"We have now identified 5,000 characteristics for identifying faces," Thornton stated, "and we apply that with a set of rules applied 200,000 times in every frame of video to determine if those characteristics are present. These include areas of dark above light, for example, which would be the eyes above cheeks. We can do this at 30 frames per second."
Not only can the software detect and recognize, but it can index as well, so that an operator can search a database of video with specific queries. There are numerous areas of development, the technical details of which can be explored on the MERL website. One of those areas of development includes audio event detection in association with video. A system is already in use where traffic accidents are being recorded based on sound detection.
Further development continues in areas like computation camera techniques that help with object tracking, which allows software to analyze the path of an object or a person and whether it's in keeping with the norm for that area. This and other advanced techniques continue to pour out of the labs in Cambridge and filter through the various product divisions such as Mitsubishi Digital Electronics America, the U.S. division for digital security products.