KGIndex

India --:--:-- --

All work

08 · Computer vision

Video-Tracer

Detect, track and count what moves through a doorway, with zones that fire events.

When
2026
Role
Author
Built with
YOLOv8, ByteTrack, BoT-SORT, OpenCV, Python
Source
KrishG7/Video-Tracer ↗

See it

A street scene with tracked people and a bus, each boxed with an ID, and red regions of interest
Detect, track, triggerEvery person and vehicle gets a class, a confidence and a persistent ID. The red boxes are regions of interest: when a tracked object enters one, an event is logged with a timestamp, here “#4 ENTERED entrance_zone”. Frame from a demo clip made from a public sample photo.

The problem

Hours of video contain a few seconds that matter. The useful question is not what is in the frame but who crossed which line, and when.

Numbers, with dates

  • About 55 FPSdetection plus tracking on an Apple M3 GPU with the small pretrained modelmeasured 28 Oct 2026
  • 0.575 mAP50for the pretrained detector on a 300-image subset of COCO val (2,147 boxes); person AP50 is 0.753run 28 Oct 2026
  • 33.2% MOTAtracking on 7 MOT16 training sequences (5,316 frames), with precision 91.5%run 28 Oct 2026

How it works

  1. 01 · detect

    Find the objects

    A YOLO detector boxes people, vehicles and other COCO classes in every frame.

  2. 02 · track

    Keep their identity

    A tracker links boxes across frames, so a person is one ID, not a hundred detections.

  3. 03 · zones

    Draw the lines that matter

    Zones and tripwires fire when a track enters or crosses them.

  4. 04 · events

    Export what happened

    Each trigger becomes a timestamped event in a JSON file.

What I built

  • A pipeline that detects with YOLO, tracks with ByteTrack or BoT-SORT, and keeps an ID on each object across frames.
  • Regions of interest and tripwires that turn tracks into timestamped events, exported as JSON.
  • Tools to capture, extract and label footage for fine-tuning a detector on doors.