Underground vehicles cannot rely on satellite positioning, but a useful location estimate can unlock both productivity and autonomy. Cameras offer a lower-cost, lower-data alternative to some ranging systems, provided the algorithm can cope with dust, low light and repetitive surfaces.
Adam Neethling’s approach combines deep-learning feature extraction with an uncertainty-driven method for correcting accumulated drift. Rather than treating every image match as equally reliable, the pipeline carries information about confidence into its trajectory estimate.
The position record can support operational indicators such as cycle time, harsh braking and slow corners. It can also become an input to navigation and autonomous functions. The research was largely complete by mid-2026, with dissertation finalisation under way.