Deep Learning-Based Detection of Open Manholes

About the Project

Inspiration

Open and damaged manholes are a serious safety hazard on roads, sidewalks, and public areas. They can be difficult to notice, especially at night, during rain, or in crowded locations. We were inspired to explore how computer vision could help identify these hazards more quickly.

What We Built

We developed a Deep Learning-Based Open Manhole Detection System that analyzes images and detects open or damaged manholes using a deep learning object-detection model.

The system takes an image as input and identifies whether an open manhole is present. When detected, the system can highlight the location of the hazard in the image, making the result easier to understand and report.

How We Built It

The project follows this workflow:

1. Collect and prepare images containing open, closed, and damaged manholes.
2. Annotate the relevant objects in the images.
3. Train a deep learning object-detection model using the prepared dataset.
4. Evaluate the trained model using detection metrics.
5. Develop a simple application where users can upload an image for detection.
6. Display the detected manhole with a bounding box and confidence score.

What We Learned

Through this project, we learned about image datasets, data annotation, model training, object detection, evaluation metrics, and deploying a trained deep learning model into a simple application.

We also learned that real-world images can vary significantly because of lighting, camera angle, weather, distance, and background conditions.

Challenges

One of the main challenges was obtaining suitable images with sufficient variation. Another challenge was improving detection performance when the manhole was partially hidden, viewed from different angles, or surrounded by similar road surfaces.

We addressed these challenges through dataset preparation, image augmentation, model training, and testing with different real-world image conditions.

Future Scope

The system can be extended into a larger road-safety platform by adding GPS-based location reporting, mobile camera detection, municipal notifications, and real-time video detection. This could help authorities identify hazardous locations and prioritize maintenance.

Built With

• computer
• deep
• learning
• python
• vision
• yolo