Klein, Jonathan: Transient Non-Line-of-Sight Imaging. - Bonn, 2021. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-62075
@phdthesis{handle:20.500.11811/9066,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-62075,
author = {{Jonathan Klein}},
title = {Transient Non-Line-of-Sight Imaging},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2021,
month = may,

note = {Sight is perhaps the most important sense of the human species. But while it allows us to gain a near-instant understanding of our surrounding, it has a fundamental limitation: An object can be hidden from view if it is occluded by an obstacle such as a building or a car. To reveal it, techniques for looking around a corner are required, which became only recently available through the use of computational photography.
In most common setups, a laser is used to illuminate a diffuse wall in the visible part of the scene from where the light can bounce off towards the hidden object. From the object, the light is reflected back onto the wall in the visible part of the scene, where it can be detected by a camera. Typically, the camera is a transient imaging camera, which can temporally resolve the propagation of light through the scene when it is illuminated by a synchronized laser. This then allows the recording of temporal light profiles on the reflector wall. The diffuse reflections destroy most of the angular information of where the light was coming from, but leave the temporal offset (caused by the travel time of the photons) intact.
The measured signal is a three-dimensional transient image in which the hidden object is not directly visible. It does, however, encode information about the hidden object which can be used together with physically based models of indirect light transport to attempt a reconstruction of the hidden object. Such a reconstruction is a challenging task which becomes apparent in the limitations of today’s system. It thus remains an active field of research that receives high interest from both academia and industry due to its many potential applications.
In this thesis we address some of the main limitations to help the field of indirect vision advance into product-ready technology. Our solutions are presented as a cumulative thesis consisting of three peer-reviewed publications:
In the first publication, we present a novel approach for real-time tracking of hidden objects. So far, setups have relied on expensive hardware and required lengthy reconstruction time. We argue, that sometimes it is more important to have real-time information about the position and movement of a target than a more precise three-dimensional reconstruction that takes minutes to obtain. Furthermore, the analysis-by-synthesis scheme that we use is extensible and works with different types of hardware including non-transient intensity cameras like webcams.
In the second publication, we present a comparison and evaluation platform for the multitude of reconstruction approaches that have been published in the previous years. The results from different research groups are usually coming from different hardware, scenes, reconstruction targets and setup scales. This makes results hard to compare, for example, when two camera system have very different signal-to-noise ratios or a scene is more challenging than another. In our benchmark, we provide a unified measurement data set that allows to run different reconstruction algorithms on the same input date and also domain specific evaluation metrics to compare the reconstruction results.
In the third publication, we present a flexible calibration algorithm that does not rely on any additional hardware. In order to estimate possible light interactions in the hidden part of the scene, knowledge about the directly visible part of the scene is used. Methods for capturing three-dimensional scenes are established but the requirement of additional hardware would increase the complexity of indirect vision systems even further. Our calibration method only facilitates an additional household-grade mirror which makes it especially suitable for the stage of lab testing in which most of the current research progress happens.
In conclusion, we present a range of contributions that partake in the global efforts of making indirect vision systems available as an additional corner stone of future vision tasks.},

url = {https://hdl.handle.net/20.500.11811/9066}
}

Die folgenden Nutzungsbestimmungen sind mit dieser Ressource verbunden:

InCopyright