As part of Centeye's participation in the Harvard University Robobee project, we are trying to see just how small we can make a vision system that can control a small flying vehicle. For the Robobee project our weight budget will be on the order of 25 milligrams. The vision system for our previous helicopter hovering system weighed about 3 to 5 grams (two orders of magnitude more!) so we have a ways to go!
We recently showed that we can control the yaw and height (heave) of a helicopter using just a single sensor. This is an improvement over the eight-sensor version used previously. The above video gives an overview of the helicopter (a hacked eFlite Blade mCX2) and the vision system, along with two sample flights in my living room. Basically a human pilot (Travis Young in this video) is able to fly the helicopter around with standard control sticks (left stick = yaw and heave, right stick = swash plate servos) and, upon letting go of the sticks, the helicopter with the vision system holds yaw and heave. Note that there was no sensing in this helicopter other than vision- there was no IMU or gyro, and all sensing/image processing was performed on board the helicopter. (The laptop is for setup and diagnostics only.)
The picture below shows the vision sensor itself- the image sensor and the optics weigh about 0.2g total. Image processing was performed on another board with an Atmel AVR32 processor- that was overkill and an 8-bit device could have been used.
A bit more about optics: In 2009 we developed a technique for "printing" optics on a thin plastic sheet, using the same photoplot process used to make masks for, say, making printed circuit boards. We can print up thousands of optics on a standard letter size sheet of plastic for about $50. The simplest version is a simple pinhole, which can be cut out of the plastic and glued directly onto an image sensor chip- pretty much any clear adhesive should work.The picture below shows a close-up of a piece of printed optics next to an image sensor (the one below is a different sensor, the 125 milligram TinyTam we demonstrated last year).
The principle of the optics is quite understandable- a cross section is below. The plastic sheet has a higher index of refraction than air, thus a near hemisphere field of view of light may be focused onto a confined region of the image sensor chip. You won't grab megapixel images in this manner, but it works well for the hundreds of pixels needed for hovering systems like this.
We are actually working on a new ArduEye system, using our newer Stonyman vision chips, to allow others to hack together sensors using this type of optics. A number of variations are possible, including using slits to sense 1D motion or pinhole arrays to make a compound eye sensor. If you want more details on this optics technique, you can visit this post, or you can pull up US patent application 12/710,073 on Google Patents. (Note: We are planning to give a blanket license of the patent for use in open hardware systems.)
(Sponsor Credit: "This work was partially supported by the National Science Foundation (award # CCF-0926148). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.")