Multi Camera Multi-Object-Tracking (MC-MOT)

Install

To install, run:

1. Create a new conda env:

conda create -n mc-mot python=3.10

NOTE: If you already have one conda env, just move to the next step.

2. Install the dependencies

2.1 PyTorch

conda install pytorch torchvision pytorch-cuda=your-cuda-version -c pytorch -c nvidia

Replace the your-cuda-version text by your CUDA version. You can check that with nvidia-smi.

However, if you don't have a GPU from NVIDIA, you can use the CPU version of PyTorch. Then, you can just install with pip:

pip install torch torchvision

2.2 Other Dependencies

Then, we can just install the other dependencies with:

pip install -r requirements.txt

Running

1. We'll download the cam1.mp4 and cam4.mp4 video files from the EPFL dataset.

cam1.mp4

cam4.mp4

Store then in a folder that is easy to access.

2. Run the homography calibration:

python3 calibrate.py --video1 /path/to/cam1.mp4 --video2 /path/to/cam4.mp4 --homography-pth /path/to/save/the/homography/matrix

You can also do python3 calibrate.py --help

At the end, you'll visualize how the bounding boxes are being projected from one camera to the other one.

3. Now, run the main.py:

python3 main.py --video1 /path/to/cam1.mp4 --video2 /path/to/cam4.mp4 --homography /path/to/homography/matrix.npy

For the other parameters, run:

python3 main.py --help

Side Notes

This repository was tested with the following specs:

1. NVIDIA GeForce 3060 Laptop GPU (6Gb).
2. i7-11800H 16 Cores.
3. Ubuntu 20.04 (LTS).
4. CudaToolkit 11.8
5. NVIDIA Driver 520.61.05

You should have similar results with similar hardware as above.

If you run on a CPU, try to change the Yolo model to the smaller one.