Computer Vision

Object detection and classification models trained on real-world imagery, deployed into operational workflows where detections drive action.

Local Government - Roadside Waste Detection

A large metropolitan council (365,000 residents, waste accounting for 21% of the council budget) needed to identify illegally dumped waste across the municipality. The solution turned the council's existing fleet of rubbish trucks into mobile survey platforms.

Cameras mounted on trucks recorded video during normal collection routes. A YOLOv5 model fine-tuned on thousands of labelled images processed the footage, detecting waste items across frames. SORT (Simple Online and Realtime Tracking) maintained object identity across consecutive frames - without temporal tracking, every frame containing the same rubbish pile generates a separate detection. Geocoded detections (matching frame timestamps to GPS coordinates from the truck) were aggregated into pickup route maps directing drivers where to collect.

The challenges were specific to real-world deployment: variable lighting, partial occlusion by parked cars and vegetation, diverse waste types (mattresses, bags, furniture), and a moving camera. False-positive optimisation was critical - drivers sent to collect shadows or garden beds erodes trust in the system.

Detections below a confidence threshold were routed to a labelling system designed for council subject matter experts - not ML engineers - to review and confirm or reject. This created a data flywheel: the system generated its own training data over time as low-confidence detections were labelled by domain experts, continuously improving the model with real-world edge cases the original training set didn't cover.

The council's broader waste program, which included this system, won Outstanding Council Project at the 2023 National Waste Conference.

Additional computer vision work includes label analysis and dataset validation for a pipe inspection defect detection platform, and dataset preparation for a hazard detection system.

Stack

YOLOv5, SORT tracking algorithm, Azure ML Studio, Python.