DEPRECATED This repository is no longer maintained due to progress made on the complete FSD platform for our vehicle. please refer to:TauRD Autonomous 2019
This is basic software intended for testing the usage of yolov3-tiny for object detection in FSAE driverless.
Any notes and suggestions are welcome!
Before using please note that:
- The weights are currently trained on a minimal dataset (200 images) with only 1 class for orange cones.
- The intention is to expand to all relevent objects for FSAE driverless with a larger dataset for each class.
- Detection and tracking should improve with a larger dataset.
- I assume tracking would work better with a natively high fps video.
- This software is currently not implemented in an FSAE car and is used mainly as a "playground" for testing at the moment.
- While training the resolution was set to 608x608.
- While in use the resoution was set to 416x416.
- Increasing the resolution will improve accuracy but reduce speed,
- CPU: ~18 fps (Intel i7-4710HQ).
- GPU: ~80 (!) fps (Nvidia Geforce 970M).
Please read the code and make sure you understand it before building and using!
Setup
-
The darknet shared library was compiled with the use of CUDA, so if no gpu is used please recompile darknet
and replace libdarknet.so (re-name the resulting darknet.so file and simply replace the one used here). -
If you do have an nvidia gpu install the CUDA toolkit and cuDNN.
-
Make sure you have OpenCV installed and properly linked in your build environment.
-
Clone this repostiory
git clone --recursive https://github.com/Galzai/ORM
-
Add libdarknet.so and the cuda libraries to your library path:
export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/home/"your-username"/ORM/Debug:$LD_LIBRARY_PATH
-
Before building make sure the path to the video file in the code is correct (Can easily be modified to use a camera if you know some basic OpenCV usage).
-
Build:
cd /home/"your-username"/ORM/Debug
make
Usage
- Enter
./ORM