Real-Time Exercise Repetition Counter & Form Analyser

A computer vision system for automatic exercise analysis using real-time pose estimation. The system counts repetitions, detects postural errors, and provides instant corrective feedback for three exercises: push-ups, jumping jacks, and lunges.

Built as a final project for the Computer Vision course at Faculdade de Ciências, Universidade do Porto (2025/2026).

---

Features

Push-Up Analysis

• Repetition counting via elbow angle FSM (DOWN: θ < 90°, UP: θ > 155°)
• Multi-parameter quality validation per repetition:
  • Depth — minimum elbow angle ≤ 96° at DOWN phase
  • Lockout — full extension with angle ≥ 130° at UP phase
  • Hip sag — shoulder-hip-ankle alignment detection with two severity levels (warning and critical)
• Adaptive thresholds calibrated over first 60 frames

Jumping Jack Analysis

• Dual-criteria synchronisation detection:
  • Arm elevation — shoulder-wrist vertical angle > 130°
  • Foot separation — normalised ankle distance > 0.20
• OPEN → CLOSED transition requires both criteria simultaneously, validating bilateral motor coordination
• False positive rate < 5% with adaptive threshold calibration

Lunge Analysis

• Bilateral working leg identification in real time via hierarchical multi-criteria system:
  1. Angular difference (weight 10): |θL − θR| > 30° → lower angle = working leg
  2. Shin verticality (weight 5): minimum knee-ankle horizontal offset
  3. Knee height (weight 2): lower knee position
  4. 3D depth (weight 1): Z-coordinate tiebreaker
• Temporal confirmation over 2 consecutive frames to prevent oscillation
• Works from any camera angle (front, side, back)
• Detects insufficient depth (θknee > 110°) and knee-over-toe fault

---

Architecture

Each module (push_up.py, jumping_jack.py, lunge.py) follows a unified pipeline:

Video capture (webcam or file)
       ↓
MediaPipe Pose (33 3D landmarks)
       ↓
Joint angle & Euclidean distance computation
       ↓
Binary FSM (UP ↔ DOWN)
       ↓
Repetition quality classification
       ↓
Real-time visual feedback overlay

---

Getting started

Prerequisites

pip install -r src/requirements.txt

Run a module

# Push-up analyser (webcam)
python src/push_up.py

# Push-up analyser (video file)
python src/push_up.py --source src/pushup1.mp4

# Jumping jack analyser
python src/jumping_jack.py

# Lunge analyser
python src/lunge.py

Controls

Key	Action
P	Pause
R	Reset counter
C	Recalibrate thresholds
Q	Quit

---

Tech stack

---

Results

Tested on 720p-1080p video at 25-30 FPS:

Exercise	Key metric
Push-up	Depth and lockout validated; hip sag detected with two severity levels
Jumping Jack	Bilateral synchronisation validated; false positive rate < 5%
Lunge	Bilateral working leg identification validated on complete and incomplete movements

Limitations: occlusions above 30% degrade precision (mitigated by visibility > 0.5 threshold); ballistic movements above 60°/frame smoothed with a 5-frame temporal buffer.