| dc.description.abstracteng | The improvements in sensor technology, e.g., the development of automotive Radio Detection and
Ranging (RADAR) or Light Detection and Ranging (LIDAR), which are able to provide a higher
detail of the sensor’s environment, have introduced new opportunities but also new challenges to
target tracking. In classic target tracking, targets are assumed as points. However, this assumption
is no longer valid if targets occupy more than one sensor resolution cell, creating the need for
extended targets, modeling the shape in addition to the kinematic parameters. Different shape
models are possible and this thesis focuses on an elliptical shape, parameterized with center,
orientation, and semi-axes lengths. This parameterization can be used to model rectangles as well.
Furthermore, this thesis is concerned with multi-sensor fusion for extended targets, which can be
used to improve the target tracking by providing information gathered from different sensors or
perspectives. We also consider estimation of extended targets, i.e., to account for uncertainties, the
target is modeled by a probability density, so we need to find a so-called point estimate.
Extended target tracking provides a variety of challenges due to the spatial extent, which need
to be handled, even for basic shapes like ellipses and rectangles. Among these challenges are the
choice of the target model, e.g., how the measurements are distributed across the shape. Additional
challenges arise for sensor fusion, as it is unclear how to best consider the geometric properties
when combining two extended targets. Finally, the extent needs to be involved in the estimation.
Traditional methods often use simple uniform distributions across the shape, which do not properly
portray reality, while more complex methods require the use of optimization techniques or large
amounts of data. In addition, for traditional estimation, metrics such as the Euclidean distance
between state vectors are used. However, they might no longer be valid because they do not
consider the geometric properties of the targets’ shapes, e.g., rotating an ellipse by 180 degree
results in the same ellipse, but the Euclidean distance between them is not 0. In multi-sensor fusion,
the same holds, i.e., simply combining the corresponding elements of the state vectors can lead to
counter-intuitive fusion results.
In this work, we compare different elliptic trackers and discuss more complex measurement
distributions across the shape’s surface or contour. Furthermore, we discuss the problems which
can occur when fusing extended target estimates from different sensors and how to handle them
by providing a transformation into a special density. We then proceed to discuss how a different
metric, namely the Gaussian Wasserstein (GW) distance, can be used to improve target estimation.
We define an estimator and propose an approximation based on an extension of the square root
distance. It can be applied on the posterior densities of the aforementioned trackers to incorporate
the unique properties of ellipses in the estimation process. We also discuss how this can be applied
to rectangular targets as well. Finally, we evaluate and discuss our approaches. We show the
benefits of more complex target models in simulations and on real data and we demonstrate our
estimation and fusion approaches compared to classic methods on simulated data. | de |