Aktuelle Studien- und Abschlussarbeiten

With the increased use of Neural Network models in the Automotive industry on solving complicating issues with sophisticated architecture. It becomes difficult to implement these models directly in the vehicle’s onboard computer. With that in mind, the objective of this work is to explore different techniques and strategies to reduce the size of the state-of-the-art models used to create Virtual Sensors using techniques, such as, Pruning, quantization or Knowledge Distillation, while maintaining the performance of the original model.

Multi-task learning (MTL) is an interesting subfield of Meta-Learning in which it refers to building a single model that learns across multiple related predictive modeling tasks. Theoretically, MTL models should perform as well, if not better, as models trained on each task individually as they benefit from Implicit data augmentation, eavesdropping and better regularization. One of key advantages of MTL is that it is easier to maintain, host and deploy a single model capable to generalize for many tasks over maintaining a model for each task.

In mission-critical applications, it is necessary to comply with predefined limits (e.g. deadlines). Therefore, adjusting the quality of results (QoR) of an application can help to keep the application within these limits.

In order to support the developer in visualising and implementing such application combinations, we want to explore different ways of modelling content and processes in an application.

Motion detection and segmentation is an important task for urban scene understanding. It is used by automated vehicles to build dynamic occupancy grids to plan their navigation. While deep learning-based approaches achieve good results in motion detection and segmentation with RGB images, designing neural networks that process 3D point clouds for the same task is quite challenging.

In a large-scale urban environment, different navigation patterns can be extracted from data. These patterns are governed by numerous factors, such as social rules and the spatiotemporal context. Neuro-symbolic modeling integrates symbolic reasoning and machine-learning-based representations altogether. It represents, consequently, a promising approach to model large-scale motion patterns that are governed by rules and uncertainty at the same time.

Neuro-symbolic modeling integrates symbolic reasoning and machine-learning-based representations altogether. For the case of autonomous driving, neuro-symbolic modeling is considered as a promising approach to represent the rule based-yet-uncertain navigation process for autonomous vehicles.

The recent advances in smart city infrastructure have provided support for a higher adoption of surveillance cameras as a mainstream crime prevention measure. However, a consequent massive deployment raises concerns about privacy issues among citizens. Augmented reality (AR) offers the needed flexibility to alter the visual data captured by cameras to achieve the needed trade-off between privacy and security.

Augmented Reality (AR) technology is enabling different applications that enhance the interaction of users with the real world. In city planning, management, and showcasing, maquettes have been widely used to have a broad view of the whole urban environment in a convenient size. However, such view is in general quite limited, when considering all the contextual information that might be necessary to analyze the environment. While projections combined with explanatory descriptions can be considered a useful workaround, they are restricted in terms of the amount of information that can be displayed and the number of stakeholders that can be served at the same time.

Existent Map services, such as Google Maps and Apple Maps, are great tools to get directions and some related information to navigate through urban environments. In spite of this, when it comes to situations where specialized or specific contextual information associated to the environment is needed those services are quite limited.

Smart environments require to recognize users and their activities in an effective and unobtrusive manner. In spite of the important advances in this area, most approaches are evaluated in very small datasets that hardly represent scenarios that can be considered fully realistic.

In the event of an emergency, such as a fire, a flooding, or an earthquake, first respondents are required to asses the situation and act promptly in scenarios that typically involve obstructions. In such situations having the possibility of visualizing the objects behind the obstructions can be crucial. In spite of the recent advances in AR x-ray approaches, making them suitable to the demanding requirements of emergency scenarios is still an unresolved issue.

Privacy is essential to preserving individuals' dignity, identity, and freedom of speech. Billions of people use applications and participate in various online services daily. However, people are mostly unaware that software and service providers often monetize their personal content and data for profits or collect usage data to infer habits and preferences. As a result of many recent data breaches and data misuse incidents, users are increasingly concerned about their privacy when using applications and digital services.

An essential quality criterion of motor vehicles is their reliability. Reliability is equivalent to the likelihood that a product will operate faultlessly for a defined mileage under given operating and environmental conditions or under a given performance framework. In this context, the knowledge about the usage behavior and the resulting state of damage on public roads is a very important information. For example, this knowledge can be used to better design operating strategies and introduce predictive maintenance.

Reconstruction of 3D models based on point cloud data has been widely studied in both computer graphics and computer vision communities. However, automatic reconstruction of models that accurately represent real-world models remains an open problem.