Al Khalili, MHD Yassin (2024)
Content-Aware Adaptation Of Point Cloud Streams: A Model-based Perspective on Processing of Point Cloud Streams.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00028630
Dissertation, Erstveröffentlichung, Verlagsversion

Kurzbeschreibung (Abstract)

The utilization of point clouds, a three-dimensional (3D) data representation, has recently experienced a significant rise in popularity. Recent advancements and affordability in 3D sensor hardware have played an essential role in driving the widespread adoption of 3D point clouds across diverse domains. These domains include virtual reality, augmented reality, volumetric video, 3D sensing for robotics, smart cities, telepresence, and automated driving applications. Consequently, the availability of point clouds, characterized by millions of data points per frame, has been increasing steadily since. However, the substantial size of point cloud data presents significant challenges regarding efficient transmission, compression, and processing. Existing methods for point cloud compression tend to prioritize data quality preservation, often overlooking the practical utilization of the data. For instance, the primary concern in autonomous vehicles is machine perception tasks, such as vehicle positioning and object detection. Thus, the focus should be more on relevant objects and less, or not on other irrelevant surrounding objects. Dedicating resources to these task-specific objects conserves valuable transmission bandwidth and enhances the overall utility of the point cloud data at the recipient's end.

In this thesis, we consider the unique characteristics of point clouds, particularly the capability to extract individual objects from the original point cloud, which are then processed and streamed independently. Our first contribution improves the understanding of how various configurations, including compression-related parameters, distance, and reduced frame rate, influence the quality of point cloud objects. For our second contribution, we investigate how these configurations affect both output quality and resource demands. Examining these relationships aims to identify configurations that maximize quality while minimizing resource consumption. Based on the collected data, we generalize our contributions by building a machine learning-based model to predict the quality of a given point cloud.

To enable the adaptability of the point cloud content to changing conditions, our third contribution explores how incorporating object-related information, such as object semantics, into point cloud content streaming impacts adaptability and delivery efficiency compared to conventional methods. Through experiments, we extensively evaluate our contributions and show the significant benefits of content-aware streaming. This approach has the potential to enhance point cloud streaming by enabling dynamic content delivery that adapts to changing scenarios.

In conclusion, this thesis introduces the concept of content-aware point cloud adaptation and compares it with alternative state-of-the-art approaches. The contributions of this thesis represent an initial step towards demonstrating the feasibility of content-aware adaptation for enhancing point cloud delivery efficiency.

Frage zum Eintrag

Redaktionelle Details anzeigen