We imply Yolov11 and Deepsort_oh for object detection and object tracking (based on the former project writer did linked here), together with SlowFast model (official) for action detection on AVA movement categories.
my GPU Cuda version is 12.4
My python version is 3.12
we use the official version of SlowFast, so we have to set up the environment of SlowFast first, but follow my following instructions is OK.
1, first we download the official SlowFast from (Thanks to reference No.2)
cd yolo_slowfast
git clone https://gitee.com/qiang_sun/SlowFast.git
2, then we download the checkpoint file from official link or here and put it under root folder
we choose the SLOWFAST_32x2_R101_50_50 model (the third one)
3, we download the file "ava.json" and "SLOWFAST_32x2_R101_50_50.yaml" from here and put them into folder "yolo_slowfast/SlowFast/demo/AVA/"
You may need to change some individual settings like
"CHECKPOINT_FILE_PATH": you may change to your own route
"INPUT_VIDEO" and "OUTPUT_FILE" you may change to your own testing video route
"BUFFER_SIZE" you may also change it which determine the interval of action detection
we also provide some test videos in here
4, then we set up the environment
pip install -r requirements.txt
pip install 'git+https://github.com/facebookresearch/fvcore'
conda install av==14.2.0 -c conda-forge
pip install -U 'git+https://github.com/facebookresearch/fvcore.git' 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
git clone https://github.com/facebookresearch/detectron2 detectron2_repo
pip install -e detectron2_repo
then
export PYTHONPATH=/path/to/SlowFast:$PYTHONPATH
Here u should change the "/path/to/" part to your own route
5, at last we run command to start
python run.py --cfg SlowFast/demo/AVA/SLOWFAST_32x2_R101_50_50.yaml
fall detection:
fight detection:
SlowFast:https://github.com/facebookresearch/SlowFast/tree/main
SlowFast的辛酸复现过程:https://blog.csdn.net/normal_lk/article/details/126138119
some other problems about Deepsort_oh you can see at here