2025/01/01 by Simon Steuernagel, Marcus Baum · 1 citation
Computer Science · Engineering · #Video Surveillance and Tracking Methods #Infrared Target Detection Methodologies #Advanced Neural Network Applications
paper · doi:10.1109/tsp.2025.3574689
Extended object tracking is concerned with estimation of object properties regarding both the kinematics and extent, i.e., shape. Particular challenges arise in case the orientation of the target is varying. Existing algorithms exhibit reduced filtering quality in difficult situations. We develop a particle filter-based elliptical extended object tracker, marginalizing the orientation from the state. Monte Carlo techniques are employed for orientation estimation, and a closed-form quadratic estimator for the semi-axis lengths with given orientation is derived. Laplace’s approximation allows for an efficient closed-form computation of the marginal measurement likelihood. In order to determine a point estimate from a set of particles, the geometrical properties of the extended object state are incorporated. Extensive evaluation is carried out, demonstrating a significant improvement in accuracy compared to state-of-the-art methods. Despite the particle-based nature of the approach, large computational burdens are avoided due to the bounded nature of the sampled (one-dimensional) state.