Engineering project case study
Human Motion Prediction for Safe Autonomous Navigation
MSc research evaluating DeMo, SEAM and SHARP on Argoverse 2 and extending SHARP with Mamba state-space encoders and controlled attention ablations.
Research focus
Autonomous vehicles must anticipate how pedestrians and other road users may move before planning a safe path. My M.Sc. thesis investigates human-motion prediction methods that balance forecasting accuracy with the computational efficiency required by real autonomous systems.
The work uses the full Argoverse 2 motion-forecasting dataset and treats average displacement error, final displacement error and miss rate as complementary measures of model quality.
Models and experiments
Baseline evaluation
I evaluate DeMo, SEAM and SHARP under a consistent framework to compare accuracy, temporal modelling and suitability for deployment.
Mamba extension
I extend SHARP with selective state-space encoders after positional embedding to investigate efficient long-range temporal representation.
Attention ablations
Controlled experiments compare baseline multi-head attention, query-key normalisation, talking-heads attention and combined variants.
Trustworthy evaluation
The analysis considers robustness and failure behaviour rather than relying on a single headline metric.
Why this matters
Accurate motion prediction can give an autonomous system more time to react, but a model must also run efficiently enough for streaming use. The project therefore connects deep-learning research with the practical constraints of safe autonomous navigation.