In my bachelor thesis I demonstrate that private individuals can build a rudimentary autonomous UAV using freely available off-the-shelf hard- and software.
For the purpose of the thesis, only a rudimentary tracking algorithm is implemented. The target is identified by a colored marker, in this case an orange safety vest. In all example videos on this page, the drone is flying completly autonomous, a human pilot is only standing by for takeoff and landing and to intervene in case of emergencies.
In the video above, the camera is rigidly connected to the UAVs frame, and all image stabilization is done in software. Using a stabilized camera mount increases tracking performance and accuracy.
Shortly after the completion of the bachelor thesis, I added the ability to recognize unmarked humans in the video feed. This is done using a linear support vector machine and HOG features, which are supported by the standard OpenCV library. Applying the new tracking algorithm to the previously recorded video data yields promising results.