Violence Against Women Detection in Videos — Deep Learning Benchmark

Comparative study of deep learning architectures for automatic detection of violence against women in real-life and surveillance videos, using spatiotemporal modeling, transfer learning, and explainable AI.

This project was developed as part of the coursework for Deep Learning / Computer Vision at Esprit School of Engineering.

---

Overview

Violence against women is a major global public health and human rights concern. Automated video-based detection systems can assist law enforcement, social services, and smart city platforms in identifying violent situations in real time — enabling faster intervention and better protection.

This project benchmarks four deep learning architectures on a unified multi-source dataset to identify the most effective approach for binary video classification: Violence vs NonViolence. The models learn to recognize patterns of physical aggression, assault, and threatening behavior from raw video footage.

Each model is evaluated end-to-end — from raw video ingestion to Grad-CAM explainability — ensuring a rigorous, reproducible, and ethically grounded comparison.

---

Features

• Multi-dataset pipeline — automatic unification, cleaning, and deduplication across 5 Kaggle datasets
• Balanced 50/50 splits — stratified train / val / test (80 / 10 / 10)
• 4 model architectures compared side-by-side
• Full evaluation suite — Accuracy, F1-score, AUC-ROC, Precision-Recall curves, Confusion Matrix
• Explainable AI — Grad-CAM 3D visualizations on all models
• Training best practices — LR warmup, early stopping, label smoothing, gradient clipping, mixed precision (AMP)
• Checkpoint system — skip retraining if model already exists

---

Model Architectures

Notebook	Architecture	Strategy	Input Size	Epochs
resnet18_from_scratch.ipynb	Custom CNN + BiGRU	Trained from scratch	112×112	50
resnet18_pretrained.ipynb	ResNet18 + BiGRU	Transfer learning	224×224	24
3dcnn_gru.ipynb	3D-CNN + BiGRU	Spatiotemporal conv	112×112	30
video-swin-transformer.ipynb	Video Swin Transformer	Stage-wise fine-tuning	224×224	5+8+…

---

Directory Structure

violence-against-women-detection/
│
├── 3dcnn_gru.ipynb                  # 3D-CNN + BiGRU model
├── resnet18_from_scratch.ipynb      # Custom CNN + BiGRU (no pretraining)
├── resnet18_pretrained.ipynb        # ResNet18 pretrained + BiGRU
├── video-swin-transformer.ipynb     # Video Swin Transformer fine-tuning
│
└── README.md

---

Datasets

All models are trained on a unified dataset merged from 4 public Kaggle sources:

Dataset	Source
Real Life Violence Situations	mohamedmustafa
Video Violence Detection	magicearth25
Data by Hedi	marzoukhedi
SmartCity CCTV Violence Detection (SCVD)	toluwaniaremu

Labels are automatically mapped using keyword matching:

• Violence → violence, fight, weaponized
• NonViolence → nonviolence, nonfight, normal, nofight

---

Tech Stack

Deep Learning

• PyTorch — model training, custom datasets, DataLoaders
• TorchVision — ResNet18 pretrained weights, transforms
• pytorchvideo — Video Swin Transformer backbone
• timm — auxiliary model components

Video Processing

• OpenCV (cv2) — frame extraction, video reading
• FFmpeg — video format conversion

Evaluation & XAI

• scikit-learn — classification report, ROC-AUC, confusion matrix
• pytorch-grad-cam — Grad-CAM 3D visualizations
• Matplotlib / Seaborn — training curves, metric plots

Experiment Management

• Kaggle Notebooks — GPU P100 / T4 environment
• tqdm — progress bars
• logging — structured pipeline logs

---

Getting Started

Prerequisites

pip install torch torchvision opencv-python scikit-learn matplotlib seaborn tqdm
pip install grad-cam pytorchvideo timm einops

Running on Kaggle

1. Upload the notebooks to Kaggle
2. Add the 4 datasets as Kaggle input sources
3. Enable GPU accelerator (P100 or T4)
4. Run cells top to bottom — the pipeline auto-skips steps already completed

Force retraining

Inside each notebook, set the flags at the top:

FORCE_RETRAIN   = True   # re-run training even if checkpoint exists
FORCE_REPROCESS = True   # re-unify and clean the dataset

---

Evaluation Metrics

Each model is evaluated on the held-out test set with:

• Accuracy & F1-Score (macro)
• AUC-ROC curve
• Precision-Recall curve & Average Precision
• Confusion Matrix
• Grad-CAM heatmaps — visual explanation of model decisions

---

Acknowledgments

This project was completed under the guidance of the faculty at Esprit School of Engineering.

Datasets provided by the Kaggle community: mohamedmustafa, magicearth25, marzoukhedi, and toluwaniaremu.

---

Topics

python · deep-learning · computer-vision · violence-detection · violence-against-women · video-classification · pytorch · resnet18 · swin-transformer · grad-cam · transfer-learning · spatiotemporal · artificial-intelligence · gender-based-violence · esprit-school-of-engineering