Background. Falls among older adults cause over 45.6 million injuries annually and cost the European Union around EUR 25 billion each year. Wi-Fi Channel State Information (CSI) sensing offers a device-free, privacy-preserving approach to fall detection, but models trained on a few subjects rarely transfer to new ones — a generalisation gap that limits reliable deployment in real care environments.

Objectives. This thesis develops, implements, and evaluates the Motion-Prototype Memory Network (MPMN), a memory-augmented framework for binary Wi-Fi CSI fall detection that learns compact prototypes of the physical motion primitives shared by all falls — sudden velocity changes, loss of balance, and postural collapse — and applies retrieval-based reasoning at inference time to generalise to unseen subjects.

Methods. The MPMN is evaluated against BiLSTM and GRU baselines on FallDeWideo and UT-HAR under strict subject-disjoint protocols across ten random seeds. An ablation study examines memory sizes K ∈ {16, 32, 64}. Primary metrics are Re-call (Fall), False Alarm Rate, F2-score, and PR-AUC.

Results. On UT-HAR, the MPMN achieves Recall 0.861, FAR 0.052, F2 0.828, and PR-AUC 0.881, with statistically confirmed improvements versus both baselines (p ≤0.020, Wilcoxon). On FallDeWideo, mean differences are not significant, but theMPMN shows 12–18× lower prediction variance — a stability advantage that matters for real-world deployment. K=16 consistently performs best on both datasets.

Conclusions. These results support motion-centric prototype representations as amore stable and generalisable alternative to action-centric recurrent baselines. Cross-environment evaluation and isolation of the memory module’s contribution from the training curriculum remain priorities for future work.