AI-Enhanced Kalman Filtering for Robust Vehicle Pose Estimation and Tracking in Real-Time Traffic Videos

In autonomous driving and ADAS, precise vehicle pose estimation and tracking are crucial but remain challenging due to perception accuracy, real-time processing, and environmental variability. Current methods struggle with these issues, especially in complex scenarios like occluded or distant vehicles. To address this, a new approach fuses AI-driven algorithm with deterministic methods, combining the robustness and computational efficiency of model-based methods with the generalization capabilities of AI-based methods. This deep-learning-based Kalman Filter aim to enhance temporal consistency in monocular vehicle pose estimation and track objects with competitive real-time capability. Our AI-based approach allows the integration of multiple modalities, such as thermal or LiDAR data and improves accuracy through filtering and a learnable motion model, boosting performance across diverse driving conditions.