RNA-Seq Pan-Cancer Clustering for Tumor Identification
Unsupervised clustering of gene expression data using K-Means and DBSCAN for automatic tumor type identification from TCGA dataset.
Computer Vision · Video Analysis · Machine Learning · YOLO
PyTorch, YOLOv11, VideoMAE, XGBoost, OpenCV
★ 3 stars on GitHub LinksCode All projects
Overview
Machine Learning solution developed for the Zindi Barbados Traffic Analysis Challenge. The goal: predict traffic congestion levels 15 minutes into the future using video data from roundabout cameras.
Challenge Constraints
- Prediction Window: 15 minutes ahead using 45 minutes of historical video
- Embargo Period: 2-minute processing delay simulation
- Multi-Camera: 4 cameras monitoring Norman Niles roundabout
- Congestion Classes: Free flowing, Light delay, Moderate delay, Heavy delay
Our Approach
Feature Extraction Pipeline
1. Basic Features
- Optical flow for motion pattern analysis
- Statistical features (intensity, edge activity, brightness)
2. YOLO-based Detection
- YOLOv11 vehicle detection and tracking
- Vehicle counting and type classification
- Speed estimation and zone-based distribution
3. Deep Learning Embeddings
- EfficientNet-B0 scene embeddings
- VideoMAE spatio-temporal features
- Meta’s Perception Encoder for video understanding
Modeling Strategy
| Model Type | Models Used |
|---|---|
| Tree-Based | Random Forest, XGBoost, LightGBM, CatBoost |
| Ensemble | Soft Voting Classifier |
| Video Transformers | Fine-tuned VideoMAE |
Results
Achieved competitive ranking in the final leaderboard using our ensemble approach combining tree-based models with video transformer features.