@phdthesis{,
    author = {Naujoks, Benjamin},
    title = {Tracking dynamischer Objekte für LiDAR-Punktwolken unter Benutzung von Dirichlet-Prozessen},
    editor = {},
    booktitle = {},
    series = {},
    journal = {},
    address = {},
    publisher = {},
    edition = {},
    year = {2023},
    isbn = {},
    volume = {},
    number =  {},
    pages = {},
    url = {},
    doi = {},
    keywords = {Autonomes Fahrzeug; Multi-Objekt-Tracking; LiDAR; Echzeitverarbeitung; Punktwolke; Maschinelles Lernen; Deep-Learning; Extended-Object-Tracking; Punktwolkenverarbeitung; Dirichlet-Prozesse; Variationelle-Inferenz},
    abstract = {The perception of dynamic objects in complex scenarios is one of the basic requirements for autonomous driving. Dynamic objects are objects whose properties are time-dependent (e. g., position and speed). Some prominent examples of dynamic objects include pedestrians, cars, and cyclists. The work presented here deals with the tracking of dynamic objects, a subfield of perception in which objects are tracked over time using a uniquely assigned number and the kinematic and shape properties of the objects are estimated. The main focus of this work is the simultaneous tracking of multiple dynamic objects (aka Multi-Object-Tracking). A major challenge in this domain is data association where assigning a measurement to an object is very complex. In this work, novel recursive Multi-Object-Tracking algorithms based on Dirichlet-Processes [Ferguson,1973] are presented. They solve the data association problem in a probabilistic manner without defining the number of objects in the scene in advance. The recursive nature of the solution is presented by using two different algorithms. First, the Greedy-Dirichlet-Process-Filter uses the so-called greedy algorithm based on [Wang et al.,2011]. Secondly, the Variational-Inference-Dirichlet-Process-Filter uses a variational method called Variational-Inference [Blei et al., 2017], which is used as an inference method. In addition, a method for robust detection of significant static objects, aka landmarks, based on the Greedy-Dirichlet-Process-Filter, has been developed. Furthermore, a novel method for simultaneous estimation of kinematic and shape properties through Extended-Object-Tracking is presented in this paper. It is based on Non-Uniform-Rational-B-Splines (NURBS) [Coons, 1967], whereby NURBS are a weighted extension of B-splines. The 3D points of a Light Detection And Ranging (LiDAR) sensor are used as measurements in this work. The Velodyne HDL-64 S2 LiDAR sensor provides a dense point cloud of the environment at the sensor frequency of 10 Hz. However, due to this dense point cloud, the runtime efficiency of each algorithm is crucial to avoid losing measurements. Pre-processing steps of point clouds for significant improvement in efficiency of downstream tracking are explained as well as new concepts in the field of point cloud representation. In addition, a Bounding-Box-Fit technique is introduced for computing the enveloping box of a point cloud belonging to an object. In this work, every newly developed method is compared with the current state of the art in real-world scenarios as well as qualitatively and quantitatively investigated. It is shown that the methods developed in this work contribute in areas of runtime efficiency and accuracy of results to the robust perception of dynamic objects in LiDAR point clouds.},
    note = {},
    
    school = {Universität der Bundeswehr München},
    
}