Synthetic vision system for recognising objects in real-time.
With the advances in robotics, a robotic vision system that performs the operations of a human eye can complement the artificial intelligence that can be one of the vital cogs in the internal architecture of a robot. However, replicating the human vision system is a difficult task. There is a need for a vision system that is capable of performing the same task as that of a human vision system such as spotting and identifying objects in real-time motion.
A research team headed by Dr. Eugenio Culurciello has developed a field programmable gate array (FPGA) processor that has been specifically designed for artificial vision. It is a specialised device, and effectively brings supercomputing power to synthetic vision. It operates about 100 times faster than a laptop computer.
This system can be trained to recognise any object in real-time, either from scratch or in an unsupervised manner. Vision is taught with the help of convolutional neural networks or ConvNets – multistage neural networks that can model the way the brains visual processing area creates invariance to size and position to identify the objects.
The researchers have developed a general-purpose system that can be programmed like a standard PC, based on a runtime reconfigurable 2D grid of computing elements with reconfiguration capabilities, somewhat similar to those of an FPGA. The major difference is that the reconfiguration can be done at runtime, allowing very diverse computations to be performed on the grid. The FPGAs are custom-made for this system and are superior to any CPU (central processing unit) or GPU (graphics processing unit). The FPGA that is used in this system is substantially more efficient. According to Dr. Eugenio Culurciello, custom designed FPGAs will always outperform general purpose ICs for specific tasks since custom-made FPGAs require only about 10 watts, and within a few years will have the capability to dissipate just 1 watt.
The heart of the system is the FPGA. Due to the power of the hardware, it is capable of analysing full-motion video in real-time. A robot equipped with this system could recognise streets, vehicles, animals, trees, and people. This technology can directly benefit any robot, be it a drone, a UAV (unmanned aerial vehicle), a driverless car, or even a humanoid robot. This system can complement a human operator, thereby reducing the need for human monitoring. This system could simultaneously monitor multiple video streams, looking for certain objects or behaviours.
The ultimate aim is to create a cognitive system that can take actions based on what the system sees. However, that would require some form of thought or intelligence. According to Dr. Culurciello, driverless cars and the first general purpose domestic robots may become available in the next decade. Synthetic vision will be standard on robots, and will be able to process text, speech, or vision in real-time. Frost & Sullivan believes a system like this will form the backbone of robotic systems in the future.
For more information contact Christie Cronje, Frost & Sullivan, +27 (0)21 680 3296, christie.cronje@frost.com, www.frost.com
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version