Detecting Obstacles using a Single Camera for Micro Air Vehicles

     Research laboratories continuously work to find new algorithms and instruments that will enable Micro Air Vehicles (MAVs) to detect obstacles when they are flying autonomously. Since the payload capacity for these aerial vehicles are limited, carrying depth sensors or mounting a stereo vision system is not the most efficient way to detect obstacles. A single camera is lightweight and gives very important information about the device's surroundings. This is why some scientists are trying to find algorithms to detect obstacles using a single camera.
     Recently scientists have tried using Optical Flow to detect obstacles and avoid them. Optical flow is an algorithm that finds the changes in different frames. One way to calculate the flow or the change between different frames is to use edge detection to find points of interest. The computer algorithm continuously does this and tracks where the edge, or the point of interest, has traveled to in the next frame. In my video below you can see how the algorithm tracks some points of interest on a ball.

Using the same concept researchers are writing algorithms to track if the object of interest is moving towards the MAV relative to MAV's camera. If this occurs than the MAV moves towards the side where optical flow occurs less rapidly. In a masters project at Land University Cognitive Science, Emil Gunnarson showed his algorithm to avoid obstacles using optical flow. You can see how the algorithm performs in the video below.

One of the biggest disadvantages of this algorithm is the fact that when the obstacle is right in front of the MAV, the algorithm cannot detect it because it is not moving to either side. The only obvious thing about an obstacle right in front of the MAV is that it is getting bigger as the MAV approaches it.
Tomoyuki Mori and Sebastian Scherer (Robotics Institute, Carnegie Mellon University) have done a study to detect obstacles which are right in front of the MAV using a single camera. Their algorithm detects objects that are getting larger and interprets this as an incoming obstacle. Their paper is available freely on the Carnegie Mellon's website. From their initial results it seems like they were successful in determining obstacle positions and avoiding them with a Quad Copter. In the video below you can see their results.


These two algorithm's could be combined but I am not sure about how much computing power that would require. Adding a system that would turn the camera to the direction of the MAV would also make these two systems a lot better. This way the MAV could use all 6 Degrees Of freedom to move instead of just one.