Journal 12, November 14

Secret Math of Fly Eyes Could Overhaul Robot Vision

When things go right, robot vision is perhaps the most attention grabbing feat in computer science. I think that has to do with the long running fascination people have for designing human like machines and computers.

While most techniques for machine vision require massive amounts of processing power. Recent developments in studying fly vision have show that much simpler systems can also be effective. An example of current complex computer vision cited in the article was the Lucas-Kanade method for machine vision which is extremely computationally intensive having to compare individual pixel changes each time the image updates.

The fly inspired computer vision algorithm is much more optimal and works by ignoring areas that don't change in color and focuses on the changing patterns. This narrower approach allows more efficient implementation of commonly needed computer vision systems such as obstacle avoidance and detection. The algorithm is a feedback loop and creates a cascading non linear system of equations according to the researchers and is not fully understood but it works.

I think this sort of vision system might be useful for in the automotive industry for self guided vehicles once the algorithms are better understood. I think a drawback might be that some of the information from the camera is seemingly discarded by this design that might be needed for some applications and the system may have to rely on more conventional computer vision techniques anyway.

Examples of current computer vision systems in robots are Domo developed at MIT CSAIL and the Honda ASIMO robots. The Domo robot has been demonstrated on video to be able to interact with a visually complex environment for specific tasks. ASIMO is mainly a walking demo robot which basic balancing and obstacle avoidance. Both of these robots are fairly good examples of the state of the art in vision and environment interaction which is needed for human like robots. The main drawbacks are still high computational requirements of both systems with Domo using a powerful networked compute cluster of 15+ computers.

Perhaps if more algorithms similar to the fly vision algorithm could be discovered by experimentation and observation of nature faster and more efficient ways to comtrol robotic systems could be developed. Interestingly early versions of algorithm have already allowed the creating tiny self guided flying robots.

I think this development is similar to other developments in math for instance it is quicker to multiply 10 x 10 than to add 10 ten times a similar gain is made here where a new way via they fly vision algorithm of doing things has allowed the implementation of complex systems with much less powerful computers.