Communication Dans Un Congrès Année : 2025

Anomaly Detection for Trustworthy Personalized Pedestrian Dead Reckoning (PDR) Navigation

Hanyuan Fu
Valerie Renaudin

Résumé

Artificial Intelligence (AI) is widely used in the field of Pedestrian Dead Reckoning (PDR) research. The personalized PDR framework MAPIN (Mobility Adapted Pedestrian Inertial Navigation) presented in our previous work shows superior accuracy and robustness compared to a generic model. However, several factors can alter the inertial signal patterns and subsequently affect the performance of the personalized PDR models. Like any other AI-based algorithm, the MAPIN framework has an opaque decision mechanism and does not provide feedback on the reliability of its predictions. To build trustworthiness into our PDR framework, an input anomaly detection mechanism is proposed. The mechanism evaluates the dissimilarity between the current input instance and the PDR model's training population, and identifies when a PDR model's performance deteriorates as a result of such dissimilarity. A persistent and significant dissimilarity could indicate either that the model is being used incorrectly, or that the original training set does not cover the necessary diversity of signal patterns, or that the distribution of the input data has changed from the training data. In the latter two cases, it is recommended to retrain the model. In three case studies, the proposed anomaly detection mechanism is evaluated on walking data with and without intentional perturbations. The experiments show that the mechanism can effectively identify input instances that lead to degraded model performance.

Fichier non déposé

Dates et versions

hal-05125057 , version 1 (23-06-2025)

Identifiants

Citer

Hanyuan Fu, Valerie Renaudin. Anomaly Detection for Trustworthy Personalized Pedestrian Dead Reckoning (PDR) Navigation. 2025 IEEE/ION Position, Location and Navigation Symposium (PLANS), Institute of Electrical and Electronics Engineers, Apr 2025, Salt Lake City, United States. pp.44-52, ⟨10.1109/PLANS61210.2025.11028260⟩. ⟨hal-05125057⟩
78 Consultations
0 Téléchargements

Altmetric

Partager

  • More