Abstract

Pursuit Alert is a counter-surveillance system that identifies vehicles which may be persistently following a user, using automatic license plate recognition (ALPR). Designed for edge deployment on platforms like Raspberry Pi and Jetson Nano, it combines YOLO-based detection with OCR to recognize and log license plates from video input. The system calculates a risk score for each observed vehicle and operates offline for privacy. This open-source project showcases real-time analysis and user feedback via a lightweight interface.

Introduction

Vehicle-based surveillance is a growing concern in personal safety and counter-intelligence efforts. In response, Pursuit Alert aims to provide an accessible, privacy-focused solution that runs autonomously on low-power edge devices. This project originated as a high school senior thesis and has since evolved into a functional real-time ALPR system. The tool allows individuals to detect and assess recurring vehicles from camera footage and assign risk scores based on observation patterns.

Methodology

The ALPR pipeline is composed of three main stages:

1. Vehicle Detection: Using Ultralytics YOLOv9c, the system first detects vehicles in each frame.
2. Plate Localization: A YOLOv8n model trained on the LHQOW dataset locates the license plate area within the vehicle bounding boxes.
3. Character Recognition: EasyOCR (english_g2) extracts alphanumeric text from the plate region, outputting the full license string.
   
   Key features include:

• Frame-skipping to balance performance and speed.
• Logging of all plate sightings, with timestamps and media snapshots.
• Risk scoring based on the mean, median, and mode of sightings across all detected plates.

The entire stack is containerized via Docker and compatible with both amd64 and arm64 environments.

Experiments

Testing was conducted using a 4K video source and downscaled to 1080p with frame skipping set to 10. This resulted in:

• Input: 30 FPS
• Output: ~3 FPS processed
• Frame processing with detection, OCR, and logging in real time
  
  
  
  Tests were run on devices including:
• A Raspberry Pi 4 (4GB RAM)
• Nvidia Jetson Nano

Input methods included both USB camera feeds and uploaded videos. System performance was measured via user-facing CPU and RAM usage bars for transparency.

Results

Pursuit Alert was able to reliably detect, log, and score recurring vehicles:

• Multiple plate sightings triggered risk score calculation.
• Risk scores effectively flagged vehicles that appeared across multiple scenes or time periods.
• Media logs and per-plate galleries provided useful context for further review.

The system demonstrated stable operation offline with no need for internet access, making it ideal for field deployment or privacy-critical use cases.

Conclusion

Pursuit Alert proves that real-time counter-surveillance is achievable using open-source tools and affordable edge hardware. It delivers reliable vehicle recognition, insightful risk analysis, and clear visual logs — all without requiring cloud processing. While future development is currently paused, the platform remains open for contributors interested in expanding its capabilities, such as adding multi-camera support, neural model upgrades, or GPS integration for pursuit localization.