Code

The code was developed and tested in MATLAB R2017b with the following toolboxes on a PC with an NVIDIA Titan Xp GPU, 64GB of RAM and an i7 processor running Ubuntu 16.04.

Note that the GPU is needed to run FlowNet2 as the tensorflow variant of the FlowNet2 is easier to use and needs cuda to run the pre-processing code.

The release version of the code differs from the final version of the code used for experiments in the paper in the following ways:

1. Uses FlowNet2 based on Tensorflow: This is easier to setup and run.
2. Uses MATLAB code: Easier to debug and not much slower than the python code.
3. Uses a simple proportional controller: A simple proportional controller provides equivalent performance to a well-tuned PID controller.
4. Uses only Foreground tracker: The difference in performance between using only a foreground and a foreground+background tracker was minimal.

You need to setup the following things to run the code:

1. PC with Ubuntu 16.04, ROS Kinetic, OpenCV 3.3.0, CUDA 8.0, Tensorflow GPU >= 1.4, Matlab >= R2017b with the following toolboxes and a WiFi module (We use this TP Link WiFi module).
2. Bebop2 with the firmware version 4.0.6.
3. Clone and setup the bebop_autonomy from here.