People detection

Wondering how crowded is your favourite beach of Gipuzkoa (Basque Country)?

Model

Based on pix2seq, currently using ResNet-50 pretrained model, COCO object detection fine-tuned checkpoints.

Accuracy

Despite of being quite accurate out-of-the-box model, it fails sometimes. This is because the camera is too far from the people. This could be handled fine-tuning the model, but it would be a time consuming task... As a baseline, it is okay.

Installation

Clone the repository and the submodules with

$ git clone --recurse-submodules https://github.com/r3v1/crowded_beach.git

and finally, install the requirements (recommended using a virtual environment):

$ pip install -r requirements.txt

Download models

Just run the Makefile in the root:

$ make download_checkpoints

Running

Example of counting people in Kontxa beach with a threshold of 0.5, saving frames every 15 seconds:

$ python src/predict.py -b Kontxa -s -t 0.5 -d 15

Troubleshooting

ModuleNotFoundError: No module named 'pix2seq'

Try adding to path modules pix2seq and src like:

export PYTHONPATH=.:pix2seq

Similar projects

• Surfer Counter