Adam Neethling investigates how a mining vehicle can estimate its own motion using cameras where satellite positioning is unavailable. Reliable underground localisation can support safety systems, equipment-performance analysis and future autonomous functions, but low light, repetitive tunnel geometry, dust and weak visual texture make it a difficult computer-vision problem.
From baseline visual odometry to a mining-specific method
Adam’s honours research established the experimental and algorithmic baseline. A stereo-camera system recorded underground-style imagery alongside inertial measurements, while LiDAR-based SLAM supplied a reference trajectory. The study implemented feature-based visual odometry using SIFT and ORB features and evaluated estimated motion against that reference.
The early baseline produced relative trajectory errors as high as 47.8%. Rather than hiding that result, the study used it to identify the practical weaknesses that a robust solution needed to address: camera calibration, stereo-camera alignment, depth estimation, local optimisation and sensor fusion.
Master’s research
The subsequent MEng work advances the pipeline with deep-learning-based feature extraction and uncertainty-driven drift correction. The aim is to retain position accuracy as visual conditions change and small frame-to-frame errors accumulate over a vehicle’s route.
The resulting motion estimate can provide more than a map location. It can help reconstruct travel paths and cycle times, identify braking and cornering events, and provide a localisation input to collision-avoidance or autonomous systems.
Progress
The work is supervised by Prof Stephan Heyns and Dr Stephan Schmidt. By the second quarter of 2026, the research was largely complete and the dissertation was being finalised.