In our experimental results [1], we outfitted three KMel k500 quadrotors with downward facing IP cameras that were capable of streaming a 640x480 video stream at 30 frames per second. The implementation of the image processing and the computation of the kinematic inputs has been done using ROS and OpenCV. The latter library includes all the necessary functions for the image processing, whereas ROS greatly simplifies communication and synchronization issues that appear when working with multiple robots. For each quadrotor there is a dedicated node in charge of processing the images of that quadrotor, extracting and publishing the SIFT descriptors to the neighboring quadrotors. Each node then computes a relative homography from its own and its neighbor’s SIFT descriptors and uses the result to estimate the relative position between neighbors. An example of this process is shown in Figure 1 (left) while an example mosaic of the feature matching and pose estimation is shown in Figure 1 (right). The image processing and formation control computation together executes in a loop at 2Hz on this system. The kinematic commands are sent to a different computer where the low-level motion control of all the quadrotors is done using differential flatness.

The experiment shown in the video below uses this technique to control the three quadrotors from some initial conditions at various alititudes to a triangle formation on a plane parallel to the ground. Then the quadrotors form a different triangle formation before concluding in a perfect line formation. Figure 2 illustrates the three formations obtained in this experiment.