https://github.com/sukhitashvili/violence-detection

Deep learning based algorithm which is capable of detecting violence in indoor or outdoor environments: fight, fire or car crash and even more
https://github.com/sukhitashvili/violence-detection

car-accidents-detection classification-model deep-learning fire-detection violence-detection

Last synced: 3 months ago
JSON representation

Deep learning based algorithm which is capable of detecting violence in indoor or outdoor environments: fight, fire or car crash and even more

• Host: GitHub
• URL: https://github.com/sukhitashvili/violence-detection
• Owner: sukhitashvili
• Created: 2021-10-03T20:36:49.000Z (over 4 years ago)
• Default Branch: main
• Last Pushed: 2024-12-31T12:07:00.000Z (over 1 year ago)
• Last Synced: 2025-05-20T00:06:42.624Z (about 1 year ago)
• Topics: car-accidents-detection, classification-model, deep-learning, fire-detection, violence-detection
• Language: Jupyter Notebook
• Homepage:
• Size: 46.2 MB
• Stars: 80
• Watchers: 1
• Forks: 24
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Table of Contents

[Introduction](#introduction)

[How to Run](#howtorun)

[Resutls](#results)

[Further Work](#work)



# Introduction

This repo presents code for Deep Learning based algorithm for

**detecting violence** in indoor or outdoor environments. The algorithm can

detect following scenarios with high accuracy: fight, fire, car crash and even

more.

To detect other scenarios you have to add **descriptive text label** of a

scenario in `settings.yaml` file under `labels` key. At this moment model can

detect 16`+1` scenarios, where one is default `Unknown` label. You can change,

add or remove labels according to your use case. The model is trained on wide

variety of data. The task for the model at training was to predict similar

vectors for image and text that describes well a scene on the image. Thus model

can generalize well on other scenarios too if you provide proper textual

information about a scene of interest.



# How to Run

First install requirements:

`pip install -r requirements.txt`

To test the model you can either run:

`python run.py --image-path ./data/7.jpg`

Or you can test it through web app:

`streamlit run app.py`

Or you can see the example code in `tutorial.ipynb` jupyter notebook

Or incorporate this model in your project using this code:

```python

from model import Model

import cv2

model = Model()

image = cv2.imread('./your_image.jpg')

image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

label = model.predict(image=image)['label']

print('Image label is: ', label)

```



# Results

Below are the resulting videos and images. I used the model to make predictions

on each frame of the videos and print model's predictions on the left side of

frame of saved videos. In case of images, titles are model's predictions. You

can find code that produces that result in `tutorial.ipynb` jupyter notebook.

![Result video](./results/output_fire.gif)

![Result video](./results/output_fight.gif)

### Result Images

![Result image](./results/3.jpg)

![Result image](./results/9.jpg)

![Result image](./results/2.jpg)

![Result image](./results/4.jpg)

![Result image](./results/10.jpg)

![Result image](./results/7.jpg)

![Result image](./results/0.jpg)



# Further Work

Needs further work for enhancements like: Batch processing support for speedup, return of

multiple suggestions, threshold fine-tuning for specific data, ect.