Calculating the speed of vehicles from video images using machine learning

This work uses darknet and YOLO for the object detection. YOLO is implemented using AlexayAB's darknet repository: link

Requirements

• Cuda 10
• cuDNN 7 - possibly not necessary but everything was ran with cuDNN so cannot say for sure
• OpenCV 4
• CMake 3.8
• Python3 - and the tqdm module (pip3 install tqdm)

---

Instructions

Object Detection

1. Download AlexayAB's darknet repository.

git clone https://github.com/AlexeyAB/darknet.git

2. Make the repository.

Set the following options in the Makefile

GPU=1
CUDNN=1
OPENCV=1
LIBSO=1

Then inside the darknet root directory run

mkdir build-release
cd build-release
cmake ..
make
make install

Alternatively just running make instead may work

3. Test the build.

To test the build first download a pre-trained weight file

wget https://pjreddie.com/media/files/yolov3.weights

Then run

./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg

Check the commandline output and the output predictions.jpg file to ensure everything ran as expected.

4. Place custom scripts in the darknet/scripts folder.

Place the scripts/get_dataset.py and the scripts/create_configs.py in the darknet/scripts folder

5. Create the dataset, test, training and config files.

Run this from the darknet root dir

python3 get_dataset.py

Follow the instructions and enter the path to the raw dataset, the proportion of testing and validation sets.
(Recommended: Training - 7, Test - 2, Validation - 1) The script will do the following:

• Create a training, test and validation folder in darknet/data/Dataset
• Flatten the category labels to be of only one singular class, and remove the cyclist and pedestrain classes.
• Create all of the necessary config files: darknet/cfg/yolov3-custom.cfg, darknet/cfg/custom.data, darknet/cfg/classes.names

5. Start training.

First download the darknet53.conv.74 weight file

wget https://pjreddie.com/media/files/darknet53.conv.74

Then start the training using the following command

./darknet detector train cfg/custom.data cfg/yolov3-custom.cfg darknet53.conv.74 -map

5. Test the model.

Test the model by running

./darknet detector test cfg/custom.data cfg/yolov3-custom.cfg backup/<weightfile>.weights

The weight files are saved in the darknet/backup folder

Camera Calibration

Place the /scripts/calibrate.py file in the darknet/scripts folder and run

python3 calibrate.py -i <calibration_image_file>

Tracking

See Vehicle_Speed_Yolo/src/yolo_consol_dll.cpp file.

Speed measurements

See Vehicle_Speed_Yolo/src/yolo_consol_dll.cpp file.

---

Resources

These instructions were written up post completion of the project and there may be some missing steps that were forgotten about.
For any issues encountered during the training please see:

• AlexayAB/darknet/README.md
• PJReddie YOLO