GoodForm AI

GoodForm AI is an innovative concept for AI-assisted physical therapy that provides real-time movement feedback during exercise sessions. The system uses a custom-built pose classification neural network to analyze user movements and provide instant corrections to ensure proper form.

Technical Architecture

Pose Classification Model

The core of GoodForm is a fully connected neural network (MLP) that classifies human poses from landmark data. Each input contains 17 body landmarks with x, y coordinates and confidence scores (51 total values).

Preprocessing: LandmarksToEmbedding

A custom TensorFlow layer converts raw landmarks into a stable 34-value embedding: • Reshape: [51] → [17, 3] to separate each landmark • Slice: Remove confidence scores, keeping only (x, y) → [17, 2] • Normalize: Apply translation and scale invariance • Flatten: Produce final [34] embedding

Landmark Normalization

The normalization step ensures the model generalizes across different camera positions and subject distances: • Translation invariance: Center pose at (0, 0) using hip midpoint • Scale invariance: Normalize by pose size (max of torso×2.5 or max landmark distance)

MLP Classifier

Dense(128, ReLU6) → Dropout(0.5) → Dense(64, ReLU6) → Dropout(0.5) → Dense(N, Softmax)

ReLU6 caps activations at 6 for numerical stability. Dropout (50%) prevents overfitting during training.

Training Setup

• Optimizer: Adam • Loss: Categorical Cross-Entropy • Metric: Accuracy

Runtime Pipeline

1. Camera → RGB frames
2. Pose Estimator → 17 landmarks with confidence
3. LandmarksToEmbedding → 34-value normalized embedding
4. MLP Classifier → Softmax probabilities
5. Post-processing → Confidence gating + temporal smoothing
6. UI → Real-time feedback, rep counting, progress tracking

Feedback System

The app implements a state machine for intelligent feedback: • NoPose: No pose detected • Tracking: Landmarks stable, monitoring • Classified: Pose identified with confidence • Hold: Detecting hold duration for static poses • Correction: Form deviation detected, showing guidance • RepCount: Counting repetitions for dynamic exercises

The platform was designed to support both individual users and integration with healthcare providers, enabling clinicians and coaches to monitor patient progress remotely.