Pré-Publication, Document De Travail Année : 2025

Weighted Least-Squares Multi-Detection Fusion and Kalman Filter-Based Tracking for Collaborative Perception Systems

Résumé

In autonomous driving, research has been focus- ing on deep learning methods for collaborative perception to address the limitations of individual perception systems. While these algorithms achieve high accuracy, they require significant communication bandwidth and open access to each agent’s object detection model design and parameters. Such requirements limit their applicability in real-time and real- world autonomous driving scenarios due to communication con- straints and agents’ need to preserve intellectual property rights over their solutions. To overcome this challenge, we propose a novel late collaborative framework for multi-detection fusion and multi-object tracking that only requires the 2D bounding box category, size, position, and orientation in bird’s-eye-view projection plane. We evaluate our solution using a custom collaborative dataset generated by adding noise to the ground truth annotations of the nuScenes database. This setup allows us to assess the impact of detection uncertainties on fusion and tracking performance in a constrained environment, evaluating the robustness of our approach in real-life scenarios where agents may provide vary- ing levels of information quality. Our framework demonstrates more robust quantitative and qualitative results (see provided video) compared to state-of-the-art methods, validating its effectiveness in realistic, noisy scenarios.

Fichier principal
Vignette du fichier
WLS_ICNSC_2025_f.pdf (2.5 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04910986 , version 1 (24-01-2025)
hal-04910986 , version 2 (11-03-2025)
hal-04910986 , version 3 (17-11-2025)

Licence

Identifiants

  • HAL Id : hal-04910986 , version 3

Citer

Maryem Fadili, Louis Lecrosnier, Redouane Khemmar, Steve Pechberti. Weighted Least-Squares Multi-Detection Fusion and Kalman Filter-Based Tracking for Collaborative Perception Systems. 2025. ⟨hal-04910986v3⟩
483 Consultations
748 Téléchargements

Partager

  • More