Fight/Violence Detection in Videos Using 3D CNN

This repository contains a dataset and a 3D CNN (Convolutional Neural Network) model trained to detect fights/violence and non-violence in videos. The model is designed to capture temporal and spatial features to identify violent activities, making it suitable for real-time surveillance and security applications.

Dataset Overview

Dataset Classes: The dataset consists of two classes:
1. Violence/Fight: Videos where physical violence is present.
2. NoViolence/NoFight: Videos with no physical confrontations.
Data Format:
- The dataset contains videos that are labeled into the above two classes.
- These videos are preprocessed and split into frames that are fed into the 3D CNN model for training and detection.

Model

3D CNN Architecture:
- The 3D CNN model is trained to detect patterns across both spatial and temporal dimensions, making it ideal for analyzing video sequences.
- The model uses 3D convolutional layers to capture motion and action-based features, which are crucial for fight/violence detection.

Purpose

The model is developed to detect violent actions in video footage. This system can be deployed in surveillance cameras, security systems, or any environment where fight/violence detection is necessary.

Key Features:

Fight/Violence Detection:
- The 3D CNN model is trained to recognize fight/violence events in videos, differentiating them from non-violent actions.
- The model processes video sequences to make predictions, utilizing temporal changes and spatial context.

Code and Usage Instructions

Pre-requisites:

Python 3.8 or higher
TensorFlow or PyTorch (depending on the implementation)
OpenCV
FFmpeg (for video preprocessing)
Required libraries as mentioned in requirements.txt

Video Preprocessing:

Extract Frames from Video: The 3D CNN model expects the input as video frames. You can extract frames from videos using the following command:
```
ffmpeg -i <input-video> -vf fps=25 <output-frame-directory>/frame_%04d.jpg
```