Automated Garbage Detection and Alert System with Live Dashboard for Post Offices using YOLOv5
CleanVision is an AI-powered computer vision system developed during the Smart India Hackathon (SIH) to enhance cleanliness monitoring in post office environments. Leveraging the capabilities of YOLOv5, the system detects various garbage items in real-time surveillance footage, triggers alerts, and offers actionable insights via an interactive dashboard.
The project promotes hygiene, accountability, and transparency in postal divisions through intelligent automation.
YOLOv5 (You Only Look Once v5) is a state-of-the-art object detection model:
- 🔲 Grid-Based Detection: Divides image into cells to detect objects in each.
- 📦 Bounding Box Prediction: Outputs coordinates & confidence scores.
- 🏷️ Class Prediction: Identifies object types like plastic, paper, metal, etc.
- 🧹 NMS Filtering: Removes overlapping predictions.
- ✅ Final Output: Only the most confident, accurate detections are displayed.
- Model: YOLOv5-m (medium configuration)
- Preprocessing: Image resizing, normalization
- Training Pipeline:
- Annotated via Roboflow
- Finetuned using PyTorch with COCO-pretrained weights
- Augmentations: random scale, flip, brightness
- Real-Time Capable: Up to ~30 FPS on GPU
📁 Training notebooks are available in the ./ipynb training_files directory.
- Upon signing up, users download
Launch AI Model.pyfrom the dashboard. - The script:
- Accepts or captures image input
- Runs the YOLOv5 model locally
- Returns predictions and displays results
🛡️ This design ensures model execution is performed locally, avoiding real-time hosting, improving privacy and reducing infrastructure load.
- Input: Live video feed (frames uploaded to Firebase)
- Processing: YOLOv5 detection on server (Render)
- Output:
- 📬 Email alert to local post office
- 📈 Updates metrics on dashboard
Tested on 532 images with 462+ annotated garbage instances.
| Metric | Score |
|---|---|
| Precision (P) | 0.96 |
| Recall (R) | 0.95 |
| mAP@0.5 | 0.975 |
| mAP@0.5:0.95 | 0.89 |
✔️ High detection accuracy and robustness, even under complex environmental conditions.
In low-light scenarios, YOLO may underperform. To combat this, we integrated:
- Enhances brightness, contrast, and detail in dim lighting
- Uses histogram equalization + residual filters
- Boosts model performance by 27% in low-light images
To ensure privacy in surveillance footage, we implemented an additional Face Blur module:
- Detects human faces in real-time video streams
- Applies Gaussian blur to anonymize detected faces
- Preserves contextual information while protecting identity
| 🚨 Alert Email | 📈 Dashboard Overview |
|
|
Team CleanVision – Finalists at Smart India Hackathon 2024
Abhijeet, Souvik, Nidhi, Shraddha, Rahul, Arzu
This project is licensed under the MIT License.