Projects · No. 052026

UNSW capstone · Computer vision

Coach Cane

For people who are blind or have low vision, learning to use the white cane.

  1. The problem

    Learning to use a white cane takes lessons with a specialist trainer, and many visually impaired individuals live far from one.

  2. The idea

    Coach Cane brings the lesson to them. The trainee stands in front of a laptop camera, and the trainer joins over a live video call.

  3. The coach

    While they practise, the computer watches how they hold and swing the cane, and says tips out loud, like “speed up”, so nothing depends on seeing the screen.

Type
Capstone, COMP9900
Year
2026
Team
Five · I built the computer vision
For
Blind and low-vision trainees
Status
Prototype, handed to the client

Cam 01 · 30 fpsGood technique00:00:00

111target 60°body frame1601211141315242326252827
Arc
58°
25–85°
Rate
0.80
0.55–1.8 /s
Off midline
0.06
≤ 0.4 shoulders
Wrist roll
52°
≤ 110°

SpokenGood technique.

Illustration · simulated readings

My part

All of the computer vision

Everything between the webcam and the spoken tip: how Coach Cane sees, what it measures and what it says.

  1. 01

    See

    Google's MediaPipe finds 33 points on the body and 21 on each hand, 30 times a second, right in the browser.

  2. 02

    Measure

    My own geometry turns those points into angles, measured against the trainee's shoulders and hips, so it works wherever the laptop sits.

  3. 03

    Judge

    It weighs a few seconds of movement at a time, never a single frame, so one glitch can't set off a tip.

  4. 04

    Speak

    One tip at a time, the most important first. When it isn't sure, it stays quiet rather than guess.

What it checks

  • Framingthe whole trainee in view, at the right distance
  • Gripa handshake grip, palm sideways, thumb on top
  • Index fingerstraight along the cane
  • Wristflicks side to side, never rolls
  • Hand positioncentred in front of the body
  • Sweep widthabout 60°, from 11 to 1 o'clock
  • Sweep speedsteady, neither too slow nor too fast
  • Sweep evennessthe same on both sides
  • The cane itselffound in the camera image, even far away
MediaPipe Pose + MediaPipe Hands + Pose estimation + Hand tracking + Geometric feature engineering + Body-frame coordinates + Ridge detection + Ray casting + Frame differencing + Time-windowed statistics + Percentile filtering + Median filtering + Hysteresis + Rule-based decision engine + Web Speech API + Evaluation harness + Threshold calibration on client video + Unit testing (33 tests) + Debugging from raw measurements + TypeScript + Real-time in-browser processing + Canvas + React + Vite + WebRTC + WebSocket + FastAPI + PostgreSQL + Docker + 

Tools and skills: MediaPipe Pose, MediaPipe Hands, Pose estimation, Hand tracking, Geometric feature engineering, Body-frame coordinates, Ridge detection, Ray casting, Frame differencing, Time-windowed statistics, Percentile filtering, Median filtering, Hysteresis, Rule-based decision engine, Web Speech API, Evaluation harness, Threshold calibration on client video, Unit testing (33 tests), Debugging from raw measurements, TypeScript, Real-time in-browser processing, Canvas, React, Vite, WebRTC, WebSocket, FastAPI, PostgreSQL, Docker

Next · EducationUNSW →
Lekhana Kalyan Raj
RésuméLinkedInGitHub