Trajectory smoothing for multiple extended objects
Jakob Bramstång · Chalmers Publication Library (Chalmers University of Technology) · 2018
From the combination of the fact that modern sensors get better resolution, and that close range tracking applications, where the objects are close to the sensor, are becoming more and more common, the field of tracking multiple extended objects arises.The property of obtaining multiple measurements per scan violates the classical assumptions.This implies that classical tracking approaches cannot be applied directly.Object tracking can be performed in many ways, depending on the method that is utilized.In some tracking applications, and in data annotation, Bayesian smoothing is an important tool to infer as much as possible from accumulated data.This thesis presents two alternative smoothing methods for tracking objects and their spatial extension; the conditional random matrix model and the factorized random matrix model.These models are a natural extension to already existing work in the Bayesian forward filtering framework for extended objects using the so called random matrix model.The linear conditional model is compared to both a linear version and a nonlinear version of the factorized model.The performance of all three models are evaluated.The models show better robustness to missed detections and few measurements, and yield better results than the existing forward filtering approaches.The linear conditional model and linear factorized model perform very similarly to one another, and outperform the nonlinear factorized model in the case when the ground truth is linear motion.In the case of non-linear ground truth motion, the results are the opposite: the nonlinear factorized model performs better.However, overall the conditional model and the linear factorized model perform better than the nonlinear factorized model, mainly due to the additional approximations needed in the nonlinear factorized model.