🚶 Temporal Gait Occlusion Detection
Research-grade PyTorch model for frame-level occlusion detection and severity estimation from gait silhouette sequences.
ResNet-18 • Transformer Encoder • CASIA-B Dataset • Multi-Task Learning
✨ Overview
Occlusions such as bags, coats, or environmental obstacles can significantly reduce the performance of gait recognition systems.
This model addresses that challenge by learning both spatial and temporal information from gait silhouette sequences. It predicts, for every frame in a walking sequence:
| Task | Output |
|---|---|
| 🔍 Occlusion Detection | Binary classification (Occluded / Clean) |
| 📈 Occlusion Severity | Continuous score between 0.0 – 1.0 |
Instead of analyzing individual frames independently, the model reasons over an entire temporal sequence using a Transformer Encoder.
🏗️ Architecture
Input Silhouette Sequence
│
▼
ResNet-18 Feature Extractor
│
▼
Feature Projection Layer
│
▼
Transformer Encoder
│
┌────────────┴────────────┐
▼ ▼
Detection Head Severity Head
(Binary Classification) (Regression)
🚀 Key Features
- ✅ Frame-level occlusion detection
- ✅ Occlusion severity estimation
- ✅ Temporal sequence modeling
- ✅ Multi-task learning framework
- ✅ Synthetic occlusion generation during training
- ✅ Mixed Precision (AMP) training
- ✅ PyTorch implementation
📚 Training Dataset
The model is trained using the CASIA-B Gait Dataset, one of the most widely used benchmark datasets for gait recognition research.
Dataset Split
| Split | Subjects |
|---|---|
| Training | 001 – 074 |
| Validation | 075 – 099 |
| Testing | 100 – 124 |
Synthetic occlusions are generated on-the-fly during training, preserving the integrity of the original dataset while improving robustness.
⚙️ Training Configuration
| Component | Value |
|---|---|
| Framework | PyTorch |
| Backbone | ResNet-18 |
| Temporal Module | Transformer Encoder |
| Optimizer | AdamW |
| LR Scheduler | Cosine Annealing |
| Mixed Precision | AMP |
| Early Stopping | Enabled |
📊 Evaluation
The model is evaluated using standard classification and regression metrics.
Detection
- Accuracy
- Precision
- Recall
- F1 Score
Severity
- Mean Absolute Error (MAE)
- Root Mean Square Error (RMSE)
🎯 Intended Use
This model is intended for:
- Academic research
- Computer Vision
- Gait Analysis
- Human Motion Analysis
- Deep Learning experimentation
- Educational purposes