Cones Detection is an application that provides a standalone interface to the cone detection model used in a computer vision system working on an autonomous Formula Student car.
In Formula Student Driverless competitions, there are static and dynamic disciplines. The static disciplines are focused on rating the car’s design, the documentation of each of the car’s systems, and the management of the team. In dynamic disciplines, the physical and autonomous capabilities of the car are tested. In all dynamic disciplines, the car is placed on a track, marked by four types of traffic cones
There are four different dynamic disciplines; there is Acceleration, Skid-pad, Trackdrive, and Autocross. In the acceleration dynamic event, the car has to drive through a straight track as fast as possible. The acceleration discipline is designed to test the car’s speed capabilities as well as its ability to steer in a stable way to keep a straight direction. In the skid-pad dynamic event, the car has to drive in 2 an eight-shaped circuit.
The car first drives two laps on the right side of the track and then transitions and drives two laps on the left side after which it exits the track. The skidpad discipline tests the ability of the autonomous system to correctly track the stage the car is in and execute the following transitions. The trackdrive and autocross dynamic disciplines are similar to each other. They both consist of driving in a closed–loop unknown track of length up to 1 km. The difference is that in the autocross discipline, the car has to drive only a single lap, while in the trackdrive the car drives 10 laps. By measuring only a single lap time in the autocross, the focus is put on the car’s ability to drive through a track it has never seen before, while in the trackdrive discipline the focus is on the car’s ability to adapt to the track and optimize its driving throughout the 10 lap drive.
DNN for cones detection using the state-of-the-art YOLO (You Only Look Once) model. Training procedure was made on the FSOCO (Formula Student Objects in Context) dataset that is publicly available as of August 2022, and on recordings gathered on PWR Racing Team car tests, annotated using Computer Vision Annotation Tool (CVAT).
DNN model weights are available for download here: weights.pt
If you are installing from source, you will need:
- Python 3.8 or later (for Linux, Python 3.8.1+ is needed)
If you want to compile with CUDA support, install the following (note that CUDA is not supported on macOS)
- NVIDIA CUDA 11.0 or above
- NVIDIA cuDNN v7 or above
- Compiler compatible with CUDA
Note: You could refer to the cuDNN Support Matrix for cuDNN versions with the various supported CUDA, CUDA driver and NVIDIA hardware
If you want to disable CUDA support, export the environment variable USE_CUDA=0.
Other potentially useful environment variables may be found in main.py.
git clone https://github.com/lukaszmichalskii/camera-perception.git
cd camera-perception
# **** OPTIONAL: virtual environment for Python setup ****
python3 -m virtualenv venv
source venv/bin/activate
# **** END OPTIONAL ****python3 -m pip install -r build_requirements.txtDefault ultralytics installation detects compute platform for PyTorch to check system
if all dependencies are installed properly:
python3 sys_check.pyAside: You may end up with incorrect CPU based PyTorch installation caused by ultralytics does not detect CUDA like:
Ultralytics YOLOv8.0.97 🚀 Python-3.8.10 torch-1.9.0+cpu CPU Setup complete ✅ (24 CPUs, 31.2 GB RAM, 35.3/91.1 GB disk) CUDA is not available. PyTorch CPU 1.9.0 will be used.or PyTorch detects CUDA but does not utilize it:
Ultralytics YOLOv8.0.97 🚀 Python-3.8.10 torch-1.9.0+cpu CPU Setup complete ✅ (24 CPUs, 31.2 GB RAM, 35.3/91.1 GB disk) CUDA is available but PyTorch CPU 1.9.0 is used. Check environment configuration for better performance.If you have CUDA enabled just follow instruction from PyTorch Get Started and override installed PyTorch version.
Aside: If you are using CUDA but PyTorch does not detect it make sure steps from
Prerequisiteswas done correctly:Ultralytics YOLOv8.0.72 🚀 Python-3.8.10 torch-2.0.0+cu117 CUDA:0 (NVIDIA GeForce RTX 3060, 12036MiB) Setup complete ✅ (24 CPUs, 31.2 GB RAM, 35.3/91.1 GB disk) CUDA is not available but PyTorch CU 2.0.0 is installed. Check environment configuration for conflicted packages.
Cones Detection can be run in two modes:
- detect cones on MP4 video recording
- standalone cones inference from image
Example run with provided resources:
# image mode
python3 src/detect.py --image docs/yolo/inference.jpg
# video mode
python3 src/detect.py --video docs/yolo/autocross.mp4If you want to save results from run enable --save flag when running application. To specify output directory
use --output <path_to_dir> option (default save to results/ directory).
# save results from image inference
python3 src/detect.py --image docs/yolo/inference.jpg --saveAside: Saving each frame in video mode is possible but generate huge output:
python3 src/detect.py --video docs/yolo/autocross.mp4 --save --output <path_to_dir>
Other options can be set via environment variables:
| Variable | Description | Default |
|---|---|---|
| CONFIDENCE | Model detection confidence threshold | 0.7 |
| RESOLUTION | Frame or image resolution, for better performance consider smaller resolutions | 1280x720 |
Cones Detection has a GNU license, as found in the LICENSE file.