Anomaly Detection in Surveillance Systems using Deep Learning

Project Overview

This project focuses on Anomaly Detection in Surveillance Systems using a 3D Convolutional Neural Network (CNN) model built with TensorFlow and Keras. The system processes video input, detects anomalies, and can trigger an alert (e.g., playing a sound) when an anomaly is detected. The model is trained on a labeled dataset of surveillance videos to distinguish between normal and anomalous events.

Key Features

• 3D CNN Architecture: Utilizes a 3D Convolutional Neural Network for processing sequential frames in video data.
• Real-Time Anomaly Detection: Capable of detecting anomalies in real-time video streams.
• Audio Alert: Plays a sound when an anomaly is detected.
• Frame Sampling: Processes every nth frame from videos to reduce computational load.
• Padding/Truncating: Ensures consistent input size by padding or truncating frames in each video.

Technologies Used

• Python: Main programming language.
• TensorFlow & Keras: For building and training deep learning models.
• OpenCV: For video frame extraction and preprocessing.
• NumPy: For efficient numerical computations.
• Google Colab: For training and testing the model using Google Drive for storage.

Project Structure

├── data/                   # Directory for dataset (not provided)
├── models/                 # Directory for saved models
├── notebooks/              # Jupyter notebooks for experimentation
├── src/                    # Python scripts for data preprocessing, model, and evaluation
│   ├── data_preprocessing.py
│   ├── model_training.py
│   └── evaluate.py
├── README.md               # Project overview and instructions
└── requirements.txt        # List of dependencies

Installation & Setup

1. Install the required dependencies by running:

pip install -r requirements.txt

2. If you're using Google Colab, mount your Google Drive to access datasets and save models:

from google.colab import drive
drive.mount('/content/drive')

Dataset

This project uses the UCF Crime Trimed-Dataset from Kaggle, a subset of the UCF Crime dataset which contains labeled surveillance videos of normal and anomalous events. The dataset should be downloaded from Kaggle and stored in your Google Drive for use in Google Colab.

Model Architecture

The model is built using 3D Convolutional layers (Conv3D) and MaxPooling3D layers to process video data (sequences of frames). Below is the architecture breakdown:

• Input Layer: Accepts a batch of video frames with dimensions (30, 64, 64, 3).
• Convolutional Layers: Use Conv3D layers to extract spatial and temporal features from the video.
• MaxPooling Layers: Use MaxPooling3D layers to reduce the spatial dimensions.
• Fully Connected Layers: Classifies the video as normal or anomalous based on the extracted features.
• Softmax Output Layer: Provides the final classification result for each video.

Results

The model provides predictions in the form of probability scores for each class. Anomalies are detected if the model's confidence in the "anomaly" class exceeds a set threshold.

Future Improvements

• Real-Time Processing: Implement real-time anomaly detection in live video streams.
• Improved Data Augmentation: Add more sophisticated augmentation techniques to increase model robustness.
• Edge Deployment: Optimize the model for running on edge devices like Raspberry Pi for real-world surveillance applications.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgements

Special thanks to the TensorFlow and Keras communities for providing open-source libraries that made this project possible.