This project was started early in desember to get an understanding of the You Only Look Once object detection technique and implementation. As well as to get a foundational code to build other applicable computer vision programs on top of in the future.
Because of GitHub's limit on file sizes, the yolov3 configurations and weights could not be added to the repo, so they need to be downloaded from elsewhere on the internet, here for example and put in the same folder as the project.
Clone the repository
git clone https://github.com/joulebit/YOLO-Object-Detection.git
Navigate into the project folder
cd YOLO-Object-Detection
Install the dependencies
pip3 install -r requirements.txt
Add the missing yolov3 files as mentioned above, and add your videos and images in their respective folders, and then you should be able to run the commands
py .\yolo.py --image images/dashcam.jpg --output output/dashcam.jpg
py .\yolo_vid.py --input videos/car_chase.mp4 --output output/car_chase.avi
Hardware used was:
Intel(R) Core(TM) i5-4210U CPU @ 1.70ghz, with 8.00 GB of RAM
Average processing time was 2 seconds per frame.
To achieve REAL Time object detection, one would need to have an Nvidia card and use CUDA as this for example.
Click the videos below to see examples of the output
Adding custom objects as described here. But it requires a good dataset and preferably a GPU.
- You Only Look Once:Unified, Real-Time Object Detection. Research paper by Joseph Redmon and others.
- Implementation example made by Adrian Rosebrock over at pyimagesearch