YOLOv4 Drone Detection

About

This is a repository enabling a Parrot Bebop 2 drone to detect and track custom objects using a trained YOLOv4 object detection model. This repository was created in collaboration with Jcm4318 as part of our Bachelor's project where we proposed a new solution to Australia's invasive Cane Toad problem.

Note: this repo should be run on a UNIX system. If a GPU is used it must be NVIDIA and you should have CUDA, CUDNN and OPENCV set up. It is highly recommended that a GPU is used.

Prerequisites

1. Trained YOLOv4 Model. See Darknet to learn how to do this.
2. Parrot Bebop 2 drone.

Prerequisites for use with GPU:

1. CUDA
2. CUDNN
3. OPENCV

Setup

1. Install dependencies

   Run pip install -r requirements.txt

2. Fetch the contents of the darknet submodule:

   Run the following:

    > git submodule init 
    > git submodule update
   

3. Edit make_darknet.sh to configure as desired.

   Makefile Options:

   GPU: Enable the use of GPU

   OPENCV: Enable the use of OpenCV

   CUDNN: Enable the use of CUDNN

   CUDNN_HALF: Enables support for tensor cores

4. Build darknet

   Run make_darknet.sh

   If when run the following error is thrown:

    ./darknet ./darknet: error while loading shared libraries: libopencv_highgui.so.3.4: cannot open shared object file: No such file or directory
   

   please run the following commands:

    sudo /bin/bash -c 'echo "/usr/local/lib" > /etc/ld.so.conf.d/opencv.conf'
    sudo ldconfig
   

Use

Set run configuration of main.py to be inside the darknet directory

1. Put trained YOLOv4 model files into corresponding Vis/IR folder inside the models directory.
2. Define area for drone in main.py using area class. Specify rectangular geofence for drone. Drone will not leave this defined space.
3. Run drone detections by calling the area.run() methodd. Specify step size between photographs, length of flight path and flight altitude.

Filestructure

main.py: Code to be edited by the user to run drone detections.

detect.py: Code containing detecor class.

DroneCode_Vis.py: Code containing area class and drone control.

multimodal_detection.py: Code enabling IR and Visible detections to be combined.

make_darknet.sh: Bash script to configure and make darkent.

darknet: Directory containing darknet files and executable.

models: Directory where trained model .weights, .data, .names and .cfg files should be placed

data: Directory where frames containing classes will be saved, and where test images / videos should be placed.