This project sought to identify fights in surveillance footage by using low-cost methods.

Python 3.8+

NumPy
Pandas
sklean
OpenCV
Pillow
TensorFlow >= 2

This repository uses a TensorFlow 2.x compatible version of the one proposed by https://github.com/Qidian213/deep_sort_yolov3.

• Weight Files

  • Download the YOLOv4 and DeepSORT weights

    mars-small128.pb
    yolov4.weights

• Environment Setup

  • Clone and compile YOLOv4 from https://github.com/AlexeyAB/darknet
  • Copy the files darknet and libdarknet.so to this project's folder
  • Place the mars-small128.pb and yolov4.weights into a weights folder

• Dataset

  • Extract the RWF-2000 dataset into this folder
  • Alternatively edit file run_all.py to change the path used

Running the run_all.py script will create an output table

Running then the script plot.py will process output tables and create comparison graphs.

These tables can also be loaded by the Facets app.

By editing the paths at the start of files extract_movement.py and conflict_detector.py , one can first run the extract_movement.py script to create a table, registering the position of each person during the video.

By then running the conflict_detector.py script, one can see how many frames are there in the video and how many of those contain potential fight scenes. Also, the tables loaded for this script can also be loaded by the visualize.py script, which displays the video while showing the personal space of each person detected.