This thesis addresses the automated identification and localization of a time-varying number of objects in a stream of sensor data. The problem is challenging due to its combinatorial nature: If the number of objects is unknown, the number of possible object trajectories grows exponentially with the number of observations. Random finite sets are a relatively new theory that has been developed to derive at principled and efficient approximations. It is based around set-valued random variables that contain an unknown number of elements which appear in arbitrary order and are themselves random. While extensively studied in theory, random finite sets have not yet become a leading paradigm in practical computer vision and robotics applications. This thesis explores random finite sets in visual tracking applications. The first method developed in this thesis combines set-valued recursive filtering with global optimization. The problem is approached in a min-cost flow network formulation, which has become a standard inference framework for multiple object tracking due to its efficiency and optimality. A main limitation of this formulation is a restriction to unary and pairwise cost terms. This circumstance makes integration of higher-order motion models challenging. The method developed in this thesis approaches this limitation by application of a Probability Hypothesis Density filter. The Probability Hypothesis Density filter was the first practically implemented state estimator based on random finite sets. It circumvents the combinatorial nature of data association itself by propagation of an object density measure that can be computed efficiently, without maintaining explicit trajectory hypotheses. In this work, the filter recursion is used to augment measurements with an additional hidden kinematic state to be used for construction of more informed flow network cost terms, e.g., based on linear motion models. The method is evaluated on public benchmarks where a considerate improvement is achieved compared to network flow formulations that are based on static features alone, such as distance between detections and appearance similarity. A second part of this thesis focuses on the related task of detecting and tracking a single robot operator in crowded environments. Different from the conventional multiple object tracking scenario, the tracked individual can leave the scene and later reappear after a longer period of absence. Therefore, a re-identification component is required that picks up the track on reentrance. Based on random finite sets, the Bernoulli filter is an optimal Bayes filter that provides a natural representation for this type of problem. In this work, it is shown how the Bernoulli filter can be combined with a Probability Hypothesis Density filter to track operator and non-operators simultaneously. The method is evaluated on a publicly available multiple object tracking dataset as well as on custom sequences that are specific to the targeted application. Experiments show reliable tracking in crowded scenes and robust re-identification after long term occlusion. Finally, a third part of this thesis focuses on appearance modeling as an essential aspect of any method that is applied to visual object tracking scenarios. Therefore, a feature representation that is robust to pose variations and changing lighting conditions is learned offline, before the actual tracking application. This thesis proposes a joint classification and metric learning objective where a deep convolutional neural network is trained to identify the individuals in the training set. At test time, the final classification layer can be stripped from the network and appearance similarity can be queried using cosine distance in representation space. This framework represents an alternative to direct metric learning objectives that have required sophisticated pair or triplet sampling strategies in the past. The method is evaluated on two large scale person re-identification datasets where competitive results are achieved overall. In particular, the proposed method better generalizes to the test set compared to a network trained with the well-established triplet loss.

Accurate snow simulation is key to capture snow's iconic visuals. Intricate methods exist that attempt to grasp snow behaviour in a holistic manner. Computational complexity prevents them from reaching real-time performance. This thesis presents three techniques making use of the GPU that focus on the deformation of a snow surface in real-time. The approaches are examined by their ability to scale with an increasing number of deformation actors and their visual portrayal of snow deformation. The findings indicate that the approaches maintain real-time performance well into several hundred individual deformation actors. However, these approaches each have their individual restrictions handicapping the visual results. An experimental approach is to combine the techniques at reduced deformation actor count to benefit from the detailed, merged deformation pattern.

Since the invention of U-net architecture in 2015, convolutional networks based on its encoder-decoder approach significantly improved results in image analysis challenges. It has been proven that such architectures can also be successfully applied in different domains by winning numerous championships in recent years. Also, the transfer learning technique created an opportunity to push state-of-the-art benchmarks to a higher level. Using this approach is beneficial for the medical domain, as collecting datasets is generally a difficult and expensive process. In this thesis, we address the task of semantic segmentation with Deep Learning and make three main contributions and release experimental results that have practical value for medical imaging. First, we evaluate the performance of four neural network architectures on the dataset of the cervical spine MRI scans. Second, we use transfer learning from models trained on the Imagenet dataset and compare it to randomly initialized networks. Third, we evaluate models trained on the bias field corrected and raw MRI data. All code to reproduce results is publicly available online.

In this thesis, the performance of the IceCube projects photon propagation code (clsim) is optimized. The process of GPU code analysis and perfor- mance optimization is described in detail. When run on the same hard- ware, the new version achieves a speedup of about 3x over the original implementation. Comparing the unmodified code on hardware currently used by IceCube (NVIDIA GTX 1080) against the optimized version run on a recent GPU (NVIDIA A100) a speedup of about 9.23x is observed. All changes made to the code are shown and their performance impact as well as the implications for simulation accuracy are discussed individually. The approach taken for optimization is then generalized into a recipe. Programmers can use it as a guide, when approaching large and complex GPU programs. In addition, the per warp job-queue, a design pattern used for load balancing among threads in a CUDA thread block, is discussed in detail.

We present a non-linear camera pose estimator, which is able to handle a combined input of point and line feature correspondences. For three or more correspondences, the estimator works on any arbitrary number and choice of the feature type, which provides an estimation of the pose on a preferably small and flexible amount of 2D-3D correspondences. We also give an analysis of different minimization techniques, parametrizations of the pose data, and of error measurements between 2D and 3D data. These will be tested for the usage of point features, lines and the combination case. The result shows the most stable and fast working non-linear parameter set for pose estimation in model-based tracking.