All Projects
Crowd Counter icon

Crowd Counter

ai computer-vision docker
Crowd Counter screenshot
Enlarge

A self-contained AI tool for estimating crowd sizes from video or photos using YOLOv8 object detection, with automated reporting and GPS grouping built in.

The Challenge

Estimating crowd sizes from event footage is tedious and error-prone when it's done by hand. Whether you're documenting a public gathering, analyzing event attendance, or reviewing security footage, counting people frame by frame just doesn't scale. Organizations needed a way to quickly process video clips or photo sets and get defensible crowd estimates without hiring specialized computer vision expertise.

Our Approach

We built a self-contained, Docker-based tool powered by YOLOv8, a state-of-the-art object detection model. Drop video files or photos into a folder, run a single command, and get back annotated images with bounding boxes around every detected person, plus summary reports with counts, averages, and peak-crowd frames. The entire pipeline runs locally with no cloud dependencies, so footage stays private.

Key Features

  • Video Processing Mode: Extracts frames at configurable intervals (every 2nd, 3rd, 5th frame, and so on), balancing accuracy against processing speed
  • Photo Processing Mode: Analyzes still images with automatic GPS-based grouping from EXIF metadata, so photos from the same location get clustered together
  • YOLOv8 Detection: Uses the lightweight yolov8n model for fast, accurate person detection with bounding box visualization
  • Auto-Rotation: Automatically corrects portrait-orientation footage to landscape for consistent processing
  • Rich Reporting: Generates an HTML summary with per-video statistics, clickable peak-frame thumbnails, and CSV exports for further analysis

    • The Results

  • Reduced crowd estimation time from hours of manual review to minutes of automated processing
  • Provided defensible, reproducible counts backed by visual evidence (annotated frames)
  • Let non-technical users run analysis with simple Docker commands
  • Kept all processing local, with no footage uploaded to external services

Technologies Used

Python, YOLOv8 (Ultralytics), OpenCV, Docker, Pandas, EXIF/GPS extraction (piexif)