No. 17of 22WaypointHardware
Pose-based rep counting for a mirror display
Smart Mirror Workout Assistant
A webcam app that tracks your pose, recognizes five exercises with joint-angle rules, counts repetitions and gives form tips. It was meant to sit behind a two-way mirror on a Raspberry Pi. The app runs on a laptop; the mirror build is not documented, and rep-count accuracy was never measured.
- When
- Feb – Jun 2025
- Status
- Built
- Team
- Course project for AI in IoT at RV University, a team of about five.
Progress
Progress60%
The rep-counting app works on a laptop. The mirror and Raspberry Pi build is not shown working, and no accuracy was measured.
What comes next
- Record 20 sets of each exercise and report rep-count accuracy.
- Photograph or film the physical mirror.
5
exercises recognized by joint-angle rules: squat, push-up, jumping jack, lunge, bicep curl
Rule-based thresholds, not a trained model. Recognition was never evaluated.
Exercise rules in the app (exercise_predictor.py)Verified. Backed by a file or public page that has been checked.Exercise rules in the app (exercise_predictor.py) — Read in the project audit of 13 November 2026.33
body landmarks tracked in each frame
By MediaPipe Pose, a pretrained third-party model.
Pose pipeline in the appVerified. Backed by a file or public page that has been checked.Pose pipeline in the app — Read in the project audit of 13 November 2026.Not measured
rep-count accuracy
No quantitative result exists. The evidence is demo screenshots only.
Project audit of 13 November 2026Verified. Backed by a file or public page that has been checked.Project audit of 13 November 2026Exhibit 01 · Walk-through
From a camera frame to a counted rep
Step 01
A frame from the webcam
OpenCV reads the camera. Nothing is trained on your video; each frame is only measured.
Step 02
33 landmarks
MediaPipe Pose, a pretrained model, marks 33 points on the body: shoulders, elbows, hips, knees and the rest.
Step 03
Joint angles and rules
The app works out the angles at the joints and compares them with fixed thresholds. Those rules pick one of five exercises and count the repetitions. It is geometry, not a trained classifier, and it has never been scored against a labeled recording.
Step 04
The screen
A Streamlit page shows the exercise, the rep count, a timer and a form tip. On the planned mirror, this is what would sit behind the glass.
01
What it is
A course project for AI in IoT at RV University, built by a team of about five between February and June 2025. The idea is a mirror that coaches: a camera behind two-way glass watches you exercise, and the display behind the glass shows the exercise it sees, your rep count and form tips.
02
What was built
The working part is a webcam app. MediaPipe Pose finds 33 body landmarks in each frame, and joint-angle thresholds, written as rules rather than trained, recognize five exercises and count repetitions: squat, push-up, jumping jack, lunge and bicep curl. A Streamlit page shows the exercise, the count, a timer and tips. A second program, a PyQt6 mirror interface with voice, an offline chatbot and a start-up service for the Raspberry Pi, starts up but has no logged workout.
03
What was found
Nothing was measured. The demo screenshots show the app running and counting, push-ups at 8 repetitions in one and bicep curls at 14 in another. They do not test whether it names the right exercise or counts correctly. The parts for the mirror were bought: a Raspberry Pi 4, a camera, a display panel and two-way glass. No photo of a finished mirror is on file, so this page does not claim one was built. The only demo frames show my face, which this site does not show; pictures will follow.
04
What's next
Record 20 sets of each exercise and report how often the count is right. Then photograph or film the mirror itself.
One question
The answer is on this page. Get it right and it goes in your passport on the map.
Challenge· stamped
How does the Smart Mirror app decide which exercise you are doing?
Exercise recognition is rule-based: thresholds on joint angles from MediaPipe's 33 landmarks, not a trained model, and its accuracy was never measured. Source: the audit of the project files, 13 November 2026.
Show the answer
Joint-angle rules on 33 pose landmarks. Exercise recognition is rule-based: thresholds on joint angles from MediaPipe's 33 landmarks, not a trained model, and its accuracy was never measured. Source: the audit of the project files, 13 November 2026.
- Stack
- Python
- MediaPipe Pose
- OpenCV
- Streamlit
- PyQt6
- Raspberry Pi 4