Solution: Computer Vision-based Contamination Detection

Benefits

Increased Efficiency 

An AI-based contamination detection system can automatically identify and classify various contaminants such as timber and other organic materials, fibreglass roof bolts, plastics (e.g. signage, bollards), and tramp metal in real time with high accuracy. This enables immediate response and isolation of contaminants, maintaining a smooth and uninterrupted material flow. By reducing the need for manual inspection and preventing blockages or equipment stoppages, the system significantly improves plant throughput and operational uptime.

Improved Product Quality

By detecting and removing contaminants before they enter downstream processes, the system ensures that only clean ore proceeds to refining. This minimises the risk of organic or foreign material interference in chemical reactions, improves product purity, and enhances the overall quality of the final product.

Enhanced Safety  

Automatic detection and rejection of contaminants reduce the likelihood of crusher or chute blockages, conveyor belt damage, and metallic impact events. This not only protects equipment from costly damage but also eliminates the need for manual intervention in hazardous areas. By reducing human exposure to moving machinery and material flow, the system provides a safer, more controlled working environment while maintaining consistent process performance.

Technology

By leveraging modern object detection neural networks, a real-time contamination detection system can automatically identify and localise unwanted materials, such as tramp metal, timber and other organic materials, fibreglass (roof bolts), plastics (bollards and signage), or debris, on a running conveyor. These models operate on high-speed video feeds, detecting foreign objects frame-by-frame and providing precise bounding boxes or segmentation masks for each detected contaminant. When integrated into the plant’s control network, such detections can trigger alarms, activate rejection mechanisms, or log contamination events for downstream analysis.

Unlike traditional rule-based or color-thresholding systems, deep learning–based detectors can adapt to variations in ore texture, lighting, and object orientation. Once trained on representative datasets, they can reliably detect different contamination types and sizes in real time, even under harsh industrial conditions.

Design Considerations

Developing an effective real-time detection system requires careful balance between speed, accuracy, and robustness. The main parameters and considerations include:

• Frame Rate and Latency: The inference time per frame must stay below the video frame interval. Low latency ensures detections align with the actual position of contaminants on the belt.

• Input Resolution: Higher resolutions improve small-object detection but increase computational load. Optimal resolution depends on object size, conveyor width, and camera height.

• Model Architecture and Size: Lightweight convolutional or transformer-based models (e.g., real-time optimised backbones) are preferred for embedded GPU or edge deployment. Quantization and pruning, CUDA optimisation can further reduce latency.

• Confidence Thresholds: Setting detection confidence appropriately avoids false positives from ore texture variations while maintaining sensitivity to rare contamination events.

• Non-Maximum Suppression (NMS): Used to consolidate overlapping detections, balancing responsiveness and detection stability.

• Lighting and Imaging Conditions: Controlled illumination or infrared imaging improves detection reliability across day/night cycles and variable ore reflectance.

To maintain continuity across frames, inter-frame tracking algorithms (e.g., optical flow or feature-based tracking) associate detections over time. This allows the system to:

• Estimate object trajectories to determine whether contaminants persist or exit the field of view.

• Filter out transient false detections from dust or glare.

• Estimate belt speed–corrected object positions, enabling precise downstream actuation (e.g., rejection gate timing).

Tracking stability depends on consistent frame timing, camera calibration, and reliable bounding box overlap between frames.