Zone Analysis — RTSP Dwell-Time Person Detection with C++, YOLO ONNX and OpenCV

A C++ application that analyzes RTSP streams or video files, detects people inside user-defined polygon zones, and flags when a person stays longer than a configured dwell-time threshold. It saves alert snapshots and writes a CSV event log.

Overview

Many surveillance and analytics scenarios don't just need "is there a person?" — they need "is a person lingering in this specific area for too long?". This project addresses that by combining object detection with zone geometry and time tracking, producing actionable events (snapshots + logs) instead of raw detections.

Motivation

I built this to develop a complete, real-time computer vision pipeline in C++ rather than a notebook prototype. It exercises detection, geometric reasoning, temporal tracking, file I/O and stream handling together — the kind of end-to-end engineering required for edge deployment.

Features

• RTSP stream and local video file input
• YOLO inference via ONNX for person detection
• User-defined polygon zones
• Dwell-time tracking per person/zone
• Alert snapshot capture when the dwell threshold is exceeded
• CSV event logging for downstream analysis
• Real-time OpenCV processing pipeline

Architecture

RTSP / Video Input
        |
   OpenCV Capture
        |
  YOLO ONNX Inference  --->  Person Detections
        |
  Zone Membership Check (polygon)
        |
  Dwell-Time Tracker (per person/zone)
        |
   +----+----+
   |         |
Snapshot   CSV Event Log
 (alerts)   (timestamped)

Tech Stack

• Language: C++
• Computer Vision: OpenCV
• AI/ML: YOLO (ONNX runtime inference)
• Input: RTSP / video files
• Output: image snapshots, CSV logs
• Build: CMake

Installation

Adjust to match the actual build setup in this repository.

git clone https://github.com/sirVnK/zone-analysis.git
cd zone-analysis
mkdir build && cd build
cmake ..
cmake --build .

Dependencies (typical): OpenCV with the appropriate ONNX inference backend. Place your YOLO .onnx model where the application expects it (see config).

Usage

# Example — adapt to the actual CLI arguments
./zone-analysis --input rtsp://<camera-url> --model model.onnx --zones zones.json

Configure zones, the dwell-time threshold, and output paths via the project's configuration before running.

Demo

Placeholders — replace with real captures (no real faces or private locations).

• docs/demo.gif
• docs/screenshots/zone-detection.png
• docs/screenshots/alert-snapshot.png

Results

Benchmark planned. Real measurements (FPS, latency, detection behavior across stream types) will be documented here once collected. No fabricated numbers are included.

Roadmap

• Multi-zone and multi-stream support
• Configurable model backends
• RTSP reconnection handling
• Unit tests and CI
• Documented benchmarks on target hardware

What I Learned

• Building a real-time video pipeline in C++ with OpenCV
• Integrating ONNX-based YOLO inference into a C++ application
• Point-in-polygon geometry for zone membership
• Temporal state tracking (dwell-time) across frames
• Producing useful artifacts (snapshots, structured logs) from a CV pipeline

Security & Privacy

This repository contains no API keys, customer data, real camera footage, identifiable faces, or private deployment details. Sample inputs (if any) are synthetic or non-sensitive.

License

MIT — see LICENSE.