Case study  |  See it  |  How it works  |  Run it  |  Patent

A rep counter can tell you how much you moved, but not whether the movement was any good. FitForm watches you through an ordinary webcam, measures your joints every frame, and only counts a rep when it goes through the full range of motion. A half squat or a swinging curl doesn't add to the number.

This repository is the first working prototype, built in November 2023: a Streamlit app that counts bicep curls on each arm, shoulder presses and squats in real time. The work became a published Indian patent application with me as a co-applicant.

See it

_{The prototype app with live counters, and pose tracking through a shoulder press.}

How it works

flowchart LR
    W["Webcam frame"] --> R["BGR to RGB<br>(OpenCV)"]
    R --> P["MediaPipe Pose<br>33 body landmarks"]
    P --> A["Joint angles<br>from landmark triplets"]
    A --> S["One state machine<br>per exercise"]
    S --> C["Rep counters<br>in Streamlit"]

Loading

1. Pose. MediaPipe Pose finds 33 body landmarks in each frame, at a minimum detection and tracking confidence of 0.5.
2. Angles. For each exercise, three landmarks make a joint angle, computed with arctan2 and folded into 0 to 180 degrees.
3. State machines. Each exercise has a "down" and an "up" threshold. A rep only counts on a full cycle, so partial movements never register.

Exercise	Angle	Counts a rep when
Bicep curl, left and right	shoulder, elbow, wrist	the arm opens past 160° and then closes under 30°
Shoulder press	hip, shoulder, elbow	the arm drops to 109° or less and then rises past 169°
Squat	hip, knee, ankle	the leg is straight past 150° and then bends under 130°

Back to top

Where it went next

The later version of FitForm (2024) built on this prototype:

• Live corrections. An OpenCV overlay shows the skeleton, stage, angle and counters on the video, so form can be fixed mid-set.
• Steady under real conditions. Confidence filtering and temporal smoothing kept it at 25 FPS on a laptop CPU and cut false rep counts by 30% across 120 annotated clips, including fast and partly hidden movements.
• Session logs. Per-exercise counts are collected with Pandas for review after the workout.

The full write-up is in the case study.

Run it

Run it on a computer with a webcam. A cloud environment such as Codespaces has no camera to read from.

git clone https://github.com/AdityaSidham/FitformAPP.git
cd FitformAPP
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirement.txt
streamlit run Experimentation.py

Open the local URL Streamlit prints, press Start Detection and step back until your whole body is in frame. Press Stop Detection to release the camera. On macOS, allow camera access for your terminal the first time.

Versions are pinned on purpose. MediaPipe 1.x removed the mp.solutions API this prototype uses, so requirement.txt holds MediaPipe 0.10.21 with the matching NumPy, OpenCV and Streamlit. Use Python 3.9 to 3.12.

packages.txt lists the system libraries OpenCV needs on Debian-based hosts (libgl1).

Patent application

AI Powered Exercise Tracking and Form Correction System. Indian patent application 202421099595, published 17 January 2025, with Aditya Sidham as a co-applicant. It is a published application, not a granted patent.

Project structure

FitformAPP/
├── Experimentation.py     the Streamlit app: capture, pose, angles, counters
├── Fitform.png            header image shown in the app
├── requirement.txt        pinned Python dependencies
├── packages.txt           system packages for Debian-based hosts
└── .devcontainer/         dev container config

Back to top

Author

Aditya Sidham, applied AI, iOS and backend engineer in Los Angeles.