A Low-Power Horizon Detection Chip

(Note:  Because the project's writeup has not been published yet, I am only providing a limited information about the project)

Unmanned micro-aerial vehicles are rapidly being developed for use as a low-cost, portable, aerial surveillance platform for semi-autonomous operation.   While they are successfully achieving flight, the sensors needed for autonomous flight (in contrast to long-range navigation) are lacking.  Obstacle avoidance and flight stability remain a problem for such small vehicles with tiny weight and power budgets.  Their small size makes them susceptible to tiny wind gusts, making the speed of processing critical for stability.

While many visual motion approaches to stabilizing aerial vehicles with low-power chips are being pursued (e.g. [1, 2]), detection of the horizon is also desirable for high-altitude pitch and roll stabilization.  Several low sensor-count horizon sensing systems have been developed for assisting aircraft pilots that utilize contrast in the infrared spectrum [3] or visible light [4].  In recent years, a team from the University of Florida (UF) has demonstrated an automatic visual-horizon finding algorithm operating on a high-speed computer on the ground that receives a transmitted color video-feed from the airplane [5].  We have devised a similar algorithm that uses an optimization approach embedded in an analog VLSI vision chip to find the best visual horizon and provide a measure of confidence.  The VLSI implementation has the potential for real-time, low-power performance and operation over a wider dynamic range of image intensities.
chip layout

In this work, we have designed and tested a low-power, real-time visual horizon sensor for use in stabilizing miniature aircraft with respect to pitch and roll in moderate-to-high altitude flight.  This sensor currently incorporates a 12x12 photoreceptor array (above) and finds a best-fit horizon line based on image intensity.  The sensor includes a “confidence-level” output for a flight control system to detect poor sensing conditions.  The chip was fabricated in a commercially-available 0.5mm CMOS process and operates on less than 2.5 milliwatts with a 5V power supply.


Each pixel of the 12x12 array contains a photoreceptor (green square in the bottom right corner), circuits that compute whether the pixel is a 'sky', or 'ground' pixel, circuits that compute the average image intensity in each class, circuits that decide whether the pixel is misclassified, and circuits that compete globally to change the horizon vector.

(click on figure to see detail)  In the example to the left, an image consisting primarily of two intensity regions (bright and dark) is projected onto the chip surface.  A global horizon vector is computed by the chip and each pixel can thus reports its classification into sky or ground (top-right panel of the figure).   The binary image shown is the final result.

The bottom-left panel of the figure to the left is the endpoint of the horizon vector (which is normal to the horizon line) as it changes over time.  The line leading from the center of the image is the final vector solution.

Each each step, the total number of pixels that consider themselves to be misclassified as sky or ground is added and reported.  This can be taken as a measure of confidence of the horizon result (low numbers means high confidence.).   For perfectly linear horizons, this number could potentially reach as low as zero, but for more realistic images, even an excellent estimate for the horizon will not be able to fit a line to make all pixels "happy".
angle estimation accuracy(Click on figure to see detail) We also measured the accuracy of the horizon measurement by projecting a calibrated horizon image onto the chip and interpreting the chip's output voltages (without correction) as a horizon angle.


[1] T. Netter and N. Franceschini, "Towards UAV Nap-of-the-Earth flight using optical flow," Advances in Artificial Life, Proceedings, vol. 1674, pp. 334-338, 1999.

[2] W. E. Green, P. Y. Oh, K. Sevcik, and G. L. Barrows, "Autonomous Landing for Indoor Flying Robots Using Optic Flow," presented at ASME Intl Mech. Engr. Congress, Washington D.C., 2003.

[3] B. Taylor, C. Bil, and S. Watkins, "Horizon Sensing Attitude Stabilization: A VMC Autopilot," presented at 18th Intl. UAV Systems Conference, Bristol, UK, 2003.

[4] Futaba, "Futaba(R) PA-2: Pilot Assist Link Auto Pilot System," (http://www.futaba-rc.com/radioaccys/futm0999.html), 2004.

[5] S. M. Ettinger, M. C. Nechyba, P. G. Ifju, and M. Waszak, "Vision-guided flight stability and control for micro air vehicles," Advanced Robotics, vol. 17, pp. 617-640, 2003.