Anomaly Detection in Video Surveillance Using Spatio-Temporal Autoencoder

This project introduces a method for detecting unusual activities in video surveillance using a spatio-temporal autoencoder. The goal is to create a deep learning model that can identify anomalies in video streams, which is a key part of modern surveillance systems. The model combines Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks to analyze both the visual and time-based patterns in video data.

Key Features

• Spatio-Temporal Analysis: The model uses CNNs to process visual information and LSTMs to understand how this information changes over time.
• Anomaly Detection: By learning to reconstruct normal video sequences, the model can detect anomalies by identifying significant differences between the original and reconstructed frames.
• Testing: The model has been tested on well-known datasets like the Chinese University of Hong Kong (CUHK) Avenue dataset and the University of California San Diego (UCSD) Ped1 dataset, which include a variety of normal and abnormal events.
• Embedded Device Deployment: The model has been successfully deployed on a Raspberry Pi 4 Model B, demonstrating its potential for real-world surveillance applications.

How It Works

1. Training: The autoencoder is trained on normal video sequences. It learns to reconstruct these sequences accurately.
2. Anomaly Detection: During testing, the model tries to reconstruct new video sequences. If the reconstruction error is high, it indicates the presence of an anomaly.
3. Evaluation: The model's performance is measured using standard metrics on datasets that include both normal and abnormal events.

Deployment on Raspberry Pi 4 Model B

The model was successfully deployed on a Raspberry Pi 4 Model B, demonstrating its feasibility for real-world surveillance applications. Here’s how the deployment was achieved:

1. Model Conversion to TensorFlow Lite (TFLite):

   • The trained TensorFlow model was converted to TFLite format using post-training quantization techniques. This reduced the model size and made it suitable for deployment on resource-constrained devices like the Raspberry Pi.

2. Setting Up the Raspberry Pi:

   • The Raspberry Pi 4 Model B was configured with the necessary libraries, including TensorFlow Lite, OpenCV, and other dependencies like libatlas and libjasper.
   • A Python script was used to load the TFLite model and perform inference on video streams captured by the Raspberry Pi camera module.

3. Running the Model:

   • The Python script was executed on the Raspberry Pi, initiating the anomaly detection process.
   • The script monitored the video stream in real-time, and if anomalies were detected, it provided relevant information on the terminal or any connected display.

4. Performance on Raspberry Pi:

   • The model ran efficiently on the Raspberry Pi, demonstrating its capability to handle real-time video analysis.
   • Despite the hardware limitations, the model maintained a good balance between accuracy and inference speed, making it suitable for practical surveillance applications.

Setting Up the Raspberry Pi for TFLite Model Deployment

This section provides a step-by-step guide to set up the Raspberry Pi 4 Model B for deploying the TFLite model.

Prerequisites

• Raspberry Pi 4 Model B
• MicroSD card with Raspbian OS installed
• Raspberry Pi camera module (optional, for real-time video capture)
• Stable internet connection

Step 1: Install Required Libraries

1. Update the Raspberry Pi:

   sudo apt update
   sudo apt upgrade

2. Install Python and Pip:

   sudo apt install python3 python3-pip

3. Install TensorFlow Lite Runtime:

   pip3 install tflite-runtime

4. Install OpenCV and Other Dependencies:

   sudo apt install libopencv-dev python3-opencv
   sudo apt install libatlas-base-dev libjasper-dev

Step 2: Set Up the Camera Module (Optional)

1. Enable the Camera Interface:

   • Run sudo raspi-config and navigate to Interfacing Options > Camera to enable the camera module.

2. Test the Camera:

   raspistill -o test.jpg

   For detailed information, you can check this link from medium.com.

Step 3: Deploy the TFLite Model

1. Copy the TFLite Model to the Raspberry Pi:

   • Transfer the .tflite model file to the Raspberry Pi using SCP or a USB drive.

2. Create a Python Script for Inference:

   • Write a Python script to load the TFLite model and perform inference on video streams.

3. Run the Script:

   python3 anomaly_detection.py

Step 4: Monitor and Optimize

• Monitor the script’s output on the terminal or a connected display.
• Optimize the script for better performance by adjusting the frame rate, resolution, or model parameters.

License

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