Extended Object Tracking Using a Gaussian Process Extent Model and Scene Flow-LiDAR Fusion
Steffen Folaasen, Martin Baerveldt, Michael Ernesto López, Nicholas Dalhaug, Annette Stahl, Edmund Brekke · 2025
High-resolution sensors such as LiDAR are often employed in Extended-Object Tracking (EOT) methods to estimate the pose, velocity, and extent of the target. However, these sensors do not measure translational and rotational motion. To address this limitation, an EOT method is proposed that fuses LiDAR information with scene flow estimates. The method may consider the scene flow between LiDAR point clouds or between stereo vision images for the current and previous time steps. The target's extent is modeled using Gaussian Processes, and the resulting EOT method is tested using simulations and real-world measurements from maritime object tracking scenarios. For real-world scenarios, the scene flow vectors are calculated using the CamLiRaft architecture. The obtained results show that the fusion of scene flow estimates reduces the error in velocity estimates, which translates to better pose and extent estimates.