The second approach is more general and uses closed contours fed into a SVM. It can detect arbitrary shapes, but requires more processing power. Through our PIXHAWK middleware, the camera image can be sent to both in parallel to allow a maximum detection.
The video shows the first, highly efficient and robust approach in action:
If you have questions on the video or approach, please contact Fabian Landau.
Please note that the detection is done onboard of our coaxial and quadrotor MAVs with no external processing. The software is part of the ai_vision PIXHAWK repository and will be made available as GNU GPLv3 code.