Bridging the Bayesian Divide: Hybrid Extended Object Tracking with LEO

Introduction

The autonomous driving community has long oscillated between two epistemological poles. On one side stand classical Bayesian trackers, which encode physical priors through rigid state-space models and offer elegant uncertainty quantification, yet fracture when reality violates their assumptions. On the other side loom deep learning architectures, which adapt to complex patterns through data-driven optimization but demand dense supervision and exhibit brittle behavior under distribution shift. This tension, between prior knowledge and adaptability, defines the central engineering challenge in modern perception systems.

The paper LEO: Graph Attention Network based Hybrid Multi Sensor Extended Object Fusion and Tracking for Autonomous Driving Applications by Mayank et al. presents a compelling synthesis of these paradigms. L