NeuroWatch-AI: A Real-Time Surveillance System Based on Spiking Neural Networks

NeuroWatch-AI is an innovative real-time surveillance system that integrates Spiking Neural Networks (SNN), YOLOv8 object detection, and optical flow analysis technologies to achieve efficient detection of violent and abnormal behaviors. Developed and released on GitHub by omkorde13 on May 25, 2026, this project represents a new direction in the integration of edge computing and neuromorphic computing, featuring both high accuracy and low latency.

• Original Author/Maintainer: omkorde13
• Source Platform: GitHub
• Original Title: "NeuroWatch-AI: Real-time neuromorphic surveillance system using Spiking Neural Networks (SNNs), YOLOv8, and Optical Flow for violence and abnormal activity detection"
• Original Link: https://github.com/omkorde13/NeuroWatch-AI
• Release Date: May 25, 2026

This project is a cutting-edge real-time surveillance platform that combines biologically inspired SNNs, computer vision technologies, and deep learning algorithms to detect violent and abnormal activities in real-time video streams.

Spiking Neural Networks (SNN)

Unlike traditional artificial neural networks, SNNs simulate the working mechanism of biological neurons, transmitting information through discrete spike signals. They have the characteristics of high energy efficiency, advantages in temporal information processing, and proximity to the human visual system. In this system, SNNs are responsible for converting motion features extracted from optical flow into spike sequences and classifying violent behaviors.

YOLOv8 Human Detection

YOLOv8 is used for real-time human localization, generating bounding boxes to identify the positions of people and providing target areas for subsequent motion analysis.

Optical Flow Analysis

Extract multi-dimensional motion features: motion amplitude and direction, acceleration vectors, motion variance detection, and direction change tracking. These features are encoded into time-series data as input to the SNN.

System Processing Flow

1. Real-time video input (camera or IP camera)
2. YOLOv8 human detection
3. Optical flow motion feature extraction
4. Temporal spike encoding
5. SNN classifier to determine violent behaviors
6. Post-prediction processing to generate confidence and threat levels

The processing results are pushed in real-time to the React front-end dashboard via the FastAPI backend and WebSockets.

Technology Stack

• AI/ML Layer: Python 3.8+, PyTorch 2.0+, snnTorch, YOLOv8, OpenCV, NumPy
• Backend Services: FastAPI, Uvicorn, WebSockets
• Front-end Interface: React 18+, Vite, TailwindCSS

According to the data published by the project, NeuroWatch-AI achieves a violence detection accuracy of 76.62% on real surveillance datasets. This achievement is attributed to:

• Temporal motion pattern learning
• Intelligent encoding from optical flow to spikes
• Motion threshold filtering to reduce false alarms
• Training on real surveillance data
• Optimized SNN architecture design

The application scenarios of NeuroWatch-AI include:

• Public Safety: Real-time surveillance in crowded places such as shopping malls, stations, and campuses
• Edge Deployment: The low energy consumption of SNNs makes them suitable for edge devices
• Fast Response: Real-time detection supports second-level alarms
• Scalable Architecture: Modular design facilitates integration with existing surveillance systems

This project represents the technological evolution direction in the field of security surveillance, integrating neuromorphic computing with traditional computer vision technologies.

NeuroWatch-AI demonstrates the great potential of neuromorphic computing in practical applications. By combining SNNs with traditional CV technologies, it achieves a balance between high accuracy and low latency, providing new design ideas for intelligent surveillance systems.

For developers exploring neuromorphic computing or building real-time video analysis systems, this project provides a complete reference implementation from model training to front-end and back-end deployment.