An intelligent computer vision system designed to detect hazardous objects in children's environments before accidents happen, providing peace of mind for parents and caregivers.
*Link to 2,000+ generated images and saved YOLOv8/v10 weights and exports.
- [Full Synthetic Dataset] —https://drive.google.com/drive/folders/1l5Z_nOrOzY2A1oY6fY6Cm-P2j8FH6YBf?usp=sharing
- Daniel Gonko
- Svetlana Gavris
- Semyon Ostrovsky
Parents and childminders face a constant struggle to identify dangerous objects hidden behind toys or lying on the floor. Traditional data collection is limited due to:
- Privacy Concerns: Collecting real images from private homes is invasive.
- Ethical Barriers: Exposing children to real dangers for photography is unacceptable.
- Data Scarcity: No existing open dataset focuses specifically on household hazards in children's rooms.
We utilize Synthetic Data Generation (SDXL) to create a robust, privacy-preserving training dataset.
The system is trained to identify four critical categories:
- Choking Hazards – Coins, small batteries, and Lego parts.
- Sharp Objects – Knives and scissors.
- Electrical Hazards – Cables, adapters, and open sockets.
- Chemical Dangers – Cleaning products, detergents, and chemical bottles.
Samples of synthetic environments and individual objects created for the training pipeline.
| Synthetic Room Scene | Isolated Hazard Objects |
|---|---|
 |
 |
| SDXL generated room with hidden hazards | |
| Sample prompt for this image:"Photorealistic nursery room, sharp scissors and a kitchen knife on a table, realistic shadows" | 100+ images per class for base detection |
Trained exclusively on 2,000+ AI-generated images.
| Confusion Matrix (Synthetic) | Training Curves (Synthetic) |
|---|---|
 |
 |
Analysis: Shows the baseline performance and initial ability to recognize hazard shapes in clean environments.
| Output examples for individual objects | Output examples for room |
|---|---|
 |
 |
Trained on a combination of synthetic scenes and real-world object photos from Roboflow.
| Confusion Matrix (Hybrid) | Training Curves (Hybrid) |
|---|---|
 |
 |
| Output examples for individual objects | Output examples for room |
|---|---|
 |
 |
| Metric | Synthetic-Only Model | Hybrid Model (Final) |
|---|---|---|
| mAP50 | 0.40 | 0.90 |
| F1-Score | 0.39 | 0.90 |
Conclusion: • AI-based computer vision can effectively detect hazards in children’s rooms
• Synthetic data helps address privacy and ethical concerns, but works best when combined with real data
• Hybrid models significantly improve detection accuracy and reduce false positives
• Separating object detection from hazard classification creates a modular and explainable system
• The proposed solution has strong potential for real-world use in smart homes and child safety systems
01_generate_images.ipynb: SDXL-based room generation pipeline.Creating and marking individual...: Object isolation and auto-labeling for only syntetic train.only_syntetic_train_yolo.ipynb: Baseline YOLO training.hybrid_model.ipynb: Advanced training with mixed datasets.First/Second/Third Presentation.pptx: Interim reports on the development progress.Final Presentation.pptx: Comprehensive project summary.