AI-Based Mine Tunnel Recognition
Overview of Our Client
Our client sought to explore an alternative approach to locating workers and machinery in underground mining environments. Traditional positioning solutions often rely on radio tags, beacons, or wired infrastructure, which can require significant installation and maintenance as mine workings expand.
We developed a proof of concept (PoC) for recognizing underground locations from camera images. The solution divides mine workings into predefined cells, such as drifts, ramps, and intersections, and classifies images according to their corresponding cell identifiers.
- Industry: Mining and Natural Resources
- Project Type: AI/ML Proof of Concept
- Application: Underground Localization
Challenge
The main challenge was locating underground objects using only camera imagery, without the use of satellite positioning or tracking equipment. We broke this task down into five sub-tasks:
- Representing mine sections as separate classification classes
- Mapping cell IDs to what the camera actually sees
- Differentiating visually similar underground environments
- Building an image classification pipeline to predict cells
- Measuring recognition accuracy for each cell separately
Main Goals
We set these objectives for the PoC:
- Build a computer vision model that recognizes underground mine sections.
- Prepare labeled cave imagery for supervised model training.
- Predict the mine cell for each camera image.
- Evaluate model predictions against known cell labels.
- Lay the groundwork for testing with helmet-mounted and vehicle-mounted cameras.
Project Overview
We developed the PoC around supervised image classification. The training example used a cave image dataset organized into numbered cells, with each identifier representing a distinct location class.
The model processed an image, extracted visual features, and predicted the corresponding cell identifier. Since identifying a zone was a classification task, a cell classification head was better suited to the PoC than exact coordinate estimation: the output represented a section of the mine rather than a point on a map. This made the PoC a form of visual place recognition at the tunnel-section level.
Solution
We designed the recognition pipeline to convert underground images into discrete location predictions.
Core System Capabilities
- Image preprocessing: Standardizes images for model input.
- Visual feature extraction: Learns image characteristics that help distinguish mine cells.
- Cell classification: Maps extracted features to predefined cell identifiers.
- Location prediction: Returns the most likely cell for a given image.
- Model evaluation: Compares predicted labels with ground-truth cell identifiers to identify classification errors.
User Workflow
- Data Preparation: Organize cave images and their corresponding cell labels into training, validation, and test datasets.
- Image Processing: Resize and normalize images to meet the model’s input requirements.
- Model Training: Train the image classifier to associate visual features with numbered cells.
- Inference: Process a new image and return the predicted cell identifier.
- Evaluation: Measure classification performance using accuracy, per-class precision and recall, and a confusion matrix.
Technology Stack
To develop the underground cell recognition PoC, we selected a Python-based technology stack that supported image preprocessing, deep learning model training, and classification performance evaluation.
Programming Language
- Python
Machine Learning
- PyTorch
- Torchvision
Model Architecture
- ViT-B/16 with a cell-specific classification head
Data Processing
- NumPy
- Pillow
Model Evaluation
- scikit-learn
- Matplotlib
Development Environment
- Jupyter Notebook
Version Control
- Git
Results
The PoC established an approach for classifying underground images into predefined mine cells, providing a foundation for further investigation of camera-based localization. The main outcomes included:
- A cell-based formulation of underground location recognition
- A training workflow using cave imagery and numbered cell labels
- An image classification pipeline designed to produce cell predictions
- An evaluation methodology for assessing classification performance
- A technical foundation for exploring camera-based localization on mining equipment and worker helmets
In a production system, camera-based cell recognition would work alongside other tools. Section-level locations can feed vehicle and worker tracking, like the GPS tracking software we build for fleets and field teams, and zone-based alerts in location-based apps. They can also be combined with data from IoT devices.
Core Team
- Solution Architect: Defined the system architecture and cell recognition workflow.
- Machine Learning Engineer: Implemented model training, inference, and classification evaluation.
- Computer Vision Engineer: Developed image preprocessing and visual recognition components.
- Data Engineer: Prepared image datasets and maintained cell-label mappings.
- QA Engineer: Tested prediction consistency and analyzed classification errors.
Related Cases
- LangChain
- Python
- IoT
- Flutter
- GPS
- Flutter
- C/C++