Communication Dans Un Congrès Année : 2024

Transformer-Based State Estimation for Multi-Target Tracking: Sensitivity Analysis against Varying Kinematic Parameters and Clutter Density

Résumé

An exploration of a Transformer's behavior is proposed in the context of multi-target tracking. We investigate the behavior properties of the state predictions made by the Transformer-based method through sensitivity analysis. The experiments focus on varying kinematic parameters, clutter density, and number of objects. The Transformer-based method demonstrates consistent accuracy on the training domain and generalization capability relative to clutter density. The results show that the Transformer Tracker outperforms the Kalman Filter and Extended Kalman Filter in terms of overall accuracy. Moreover, it is capable of adapting to the training data, building data knowledge to improve prediction accuracy. The experiments highlight insufficient generalization on the number of objects. They also provide preliminary insight into the system's explainability. Finally, we discuss potential limitations and identify future research directions.
Fichier principal
Vignette du fichier
CAID2024_paper_12.pdf (401) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04937295 , version 1 (10-02-2025)

Identifiants

  • HAL Id : hal-04937295 , version 1

Citer

Valentin Sonntag, Jean-Marc Le Caillec, Alain Peres, Stéphane Devaud. Transformer-Based State Estimation for Multi-Target Tracking: Sensitivity Analysis against Varying Kinematic Parameters and Clutter Density. Conference on Artificial Intelligence for Defense, Nov 2024, Rennes, France. ⟨hal-04937295⟩
0 Consultations
0 Téléchargements

Partager

More