loading page

DeepKalPose: An Enhanced Deep-Learning Kalman Filter for Temporally Consistent Monocular Vehicle Pose Estimation
  • Leandro Di Bella,
  • Yangxintong Lyu,
  • Adrian Munteanu
Leandro Di Bella
Vrije Universiteit Brussel

Corresponding Author:[email protected]

Author Profile
Yangxintong Lyu
Vrije Universiteit Brussel
Author Profile
Adrian Munteanu
Vrije Universiteit Brussel
Author Profile

Abstract

In this paper, we introduce an innovative temporal consistency enhancement approach, which enables image-based models on video data by leveraging a deep-learning-based Kalman Filter. More specifically, we propose a novel Bi-direction Kalman filter strategy, utilizing forward and backward processing to capitalize on higher-quality pose estimations near the camera, enhancing the robustness and precision of vehicle tracking across varying distances and conditions. Then, rather than using the conventional mathematical motion model, we propose a learnable motion model, dubbed Future State Predictor, to represent the complex, non-linear motion patterns observed in vehicles. The experimental results demonstrate that our approach enhances pose accuracy and temporal consistency, which allows us to handle the challenging occluded/distant vehicles.
Submitted to Electronics Letters
24 Jan 2024Assigned to Editor
24 Jan 2024Submission Checks Completed
28 Jan 2024Reviewer(s) Assigned
06 Feb 2024Review(s) Completed, Editorial Evaluation Pending
08 Feb 2024Editorial Decision: Revise Major
09 Mar 20241st Revision Received
24 Mar 2024Review(s) Completed, Editorial Evaluation Pending
27 Mar 2024Editorial Decision: Accept