Multimedia ONline ARchiv CHemnitz
Not a member yet
4978 research outputs found
Sort by
Multimodal Sensor Fusion with Object Detection Networks for Automated Driving
Get PDFObject detection is one of the key tasks of environment perception for highly automated vehicles. To achieve a high level of performance and fault tolerance, automated vehicles are equipped with an array of different sensors to observe their environment. Perception systems for automated vehicles usually rely on Bayesian fusion methods to combine information from different sensors late in the perception pipeline in a highly abstract, low-dimensional representation. Newer research on deep learning object detection proposes fusion of information in higher-dimensional space directly in the convolutional neural networks to significantly increase performance. However, the resulting deep learning architectures violate key non-functional requirements of a real-world safety-critical perception system for a series-production vehicle, notably modularity, fault tolerance and traceability.
This dissertation presents a modular multimodal perception architecture for detecting objects using camera, lidar and radar data that is entirely based on deep learning and that was designed to respect above requirements. The presented method is applicable to any region-based, two-stage object detection architecture (such as Faster R-CNN by Ren et al.). Information is fused in the high-dimensional feature space of a convolutional neural network. The feature map of a convolutional neural network is shown to be a suitable representation in which to fuse multimodal sensor data and to be a suitable interface to combine different parts of object detection networks in a modular fashion. The implementation centers around a novel neural network architecture that learns a transformation of feature maps from one sensor modality and input space to another and can thereby map feature representations into a common feature space. It is shown how transformed feature maps from different sensors can be fused in this common feature space to increase object detection performance by up to 10% compared to the unimodal baseline networks. Feature extraction front ends of the architecture are interchangeable and different sensor modalities can be integrated with little additional training effort. Variants of the presented method are able to predict object distance from monocular camera images and detect objects from radar data.
Results are verified using a large labeled, multimodal automotive dataset created during the course of this dissertation. The processing pipeline and methodology for creating this dataset along with detailed statistics are presented as well
Berührungslose Schweißdetektion an Fahrzeuginsassen für die Automatisierung der Klimasteuerung
Get PDFDie vorliegende Arbeit beschäftigt sich mit der berührungslosen Schweißdetektion im Fahrzeug als Bestandteil der automatisierten Klimasteuerung. Für die Erkennung von
Schweiß wird eine Wärmebildkamera herangezogen, da austretender Schweiß durch Entzug der Verdampfungswärme die Haut abkühlt und daher auf dem Wärmebild zu sehen ist. Die Ergebnisse basieren auf verschiedenen Studien, in denen 20 Probanden ein sportliches Training absolvieren mussten, um ins Schwitzen zu kommen. Dabei wurde
in regelmäßigen Abständen nach dem Schweißstatus gefragt und der Kopfbereich im Wärmebild aufgenommen. Für die Auswertung wurde der Stirnbereich auf dem Wärmebild explorativ untersucht, da dieser meistens nicht von Haaren oder Kleidung bedeckt ist. Der Stirnbereich wurde dann für verschiedene Schweißzustände betrachtet und die
Temperaturdifferenz zwischen maximaler und minimaler Temperatur, sowie die festgestellte Fleckigkeit anhand einer räumlichen Frequenzanalyse analysiert. Die Ergebnisse
zeigen, dass anhand der 2D-Fouriertransformation, mit anschließender Bildung des radialen Mittelwerts aus dem Powerspektrum, die beste Unterscheidung zwischen einem
schwitzenden und nicht schwitzenden Insassen gemacht werden können. Speziell bei der Betrachtung eines Pixelbereichs von 21 Pixel x 21 Pixel auf der Stirn konnte mithilfe der
Signalentdeckungstheorie eine Treffer-Rate von 95 % unter 5 % falschen Alarmen festgestellt werden. Das bedeutet, dass es möglich ist, anhand eines Wärmebilds der Stirn das
Schwitzen bei einem Insassen im Fahrzeug mit einer Sicherheit von 95 % zu erkennen
Development and reliability quantification of a novel test set-up for measuring footwear bending stiffness
No full textSince footwear flexibility impacts functional design factors, numerous studies have investigated footwear bending stiffness. However, the various methods used to measure footwear bending stiffness have some limitations. Hence, the scope of this study was to develop and quantify the reliability of a novel test set-up for measuring footwear bending stiffness. A test set-up consisting of a hydraulic testing machine, a bending apparatus and a fixation unit was created that fulfilled the requirements specified in the initial phase of the study. The test set-up was evaluated by testing 15 different boots in three series of measurements. Bending stiffness of the boots ranged from 0.61 ± 0.03 Nm/° to 2.38 ± 0.08 Nm/°. Two-way analysis of variance test yielded that the test set-up enabled the reliable measurement of footwear bending stiffness. Relative measurement uncertainty ranged from 1.3 % to 6.1 %
Hand Sign Recognition based on Myographic Methods and Random K-Tournament Grasshopper Extreme Learner
No full textThe recognition of hand signs is important for several applications. Myographic methods, such as surface electromyography (sEMG), force myography (FMG) and
surface electrical impedance myography (sEIM) deliver interesting physiological signals, can build the basis for hand signs recognition and are subjected to investigation by a limited number of sensors at suitable positions and adequate signal processing algorithms for perspective implementation in wearable embedded systems. A database has been collected with measurements of American sign language (ASL) gestures at the forearm and wrist of more than 100 healthy persons in total.
Novel classification methods have been developed based on Extreme Learning Machine (ELM) supported by a grasshopper optimization algorithm (GOA) as a core
weight pruning process. A K-tournament selection strategy introduced to the GOA ensures its population’s diversity. The K-Tournament Grasshopper Optimization Algorithm (KTGOA) has been further improved for discrete optimization problems and implemented to select the ELM weights as a K-Tournament Grasshopper Extreme Learner (KTGEL). To improve the balance between exploration and exploitation, the balancing coefficients of the KTGEL are subjected to uniform randomization. The resulting Random K-Tournament Grasshopper Extreme Learner (RKTGEL) is a novel classifier with a simultaneously automated feature selection.
In a first stage and based on the conventional ELM method, the number of sensors and their positions have been investigated. For FMG, it has been shown that 8
sensors are suitable, for sEMG, only 2 sensors are suitable and for sEIM, 4 equidistant electrodes are suitable for measurements in the frequency range from 1 kHz to 4 kHz. In a second step, different collections of hand signs having reduced ambiguity, middle ambiguity and great ambiguity have been defined and subjected to classification by the novel algorithms. Combinations between sEMG and FMG or between sEIM and FMG reach thereby an accuracy of 100% in the cases of small and medium ambiguous signs collection even with data collected from at least 20 subjects. However, for the case of a high ambiguity, a targeted reduction of ambiguity by excluding hand signs with a high similarity is necessary. From the set of 20 gestures with a high level of ambiguity and after excluding respectively one hand sign from 6 ambiguous gesture pairs and the ambiguous sign ’Z’, 13 signs remained including letters and numbers collected from 40 subjects with 2 sEMG and 8 FMG sensors. After reduction of ambiguity, the RKTGEL reached an overall accuracy of 97%.:1 Introduction
2 Hand Gesture Recognition based on Myography
3 Extreme Learning Machine
4 Random K-Tournament Grasshopper Extreme Learner
5 Data Collection and Pre-Processing
6 Implementation of the Random K-Tournament Grasshopper Extreme Learner
7 Summary and OutlookDie Handzeichenerkennung ist für verschiedene Anwendungen wichtig. Myographische Methoden, wie die Oberflächen-Elektromyographie (sEMG), die Kraft-Myographie (FMG) und die Oberflächen-Elektrische-Impedanz-Myographie (sEIM) liefern interessante physiologische Signale, die die Grundlage für die Erkennung von Handzeichen bilden können und mit einer begrenzten Anzahl von Sensoren an geeigneten Positionen und geeigneten Signalverarbeitungsalgorithmen für eine perspektivische Implementierung in tragbare eingebettete Systeme untersucht werden müssen. Eine Datenbank mit Messungen von Gesten der amerikanischen Gebärdensprache (ASL) am Unterarm und Handgelenk von insgesamt mehr als 100 gesunden Personen wurde erhoben.
Neuartige Klassifizierungsmethoden wurden entwickelt, die auf einer Extreme Lernmaschine (ELM) basieren, unterstützt durch einen Grashüpfer-Optimierungsalgorithmus (GOA) als zentralen Prozess für das Verbindungspruning. Eine K-Tournament-Auswahlstrategie, die in den GOA eingeführt wurde, gewährleistet die Diversität seiner Population. Der K-Tournament-Grashüpfer-Optimierungsalgorithmus (KT-GOA) wurde für diskrete Optimierungsprobleme weiter verbessert und zur Auswahl der ELM-Gewichte als K-Tournament-Grashüpfer-Extrem-Lerner (KTGEL) implementiert. Um das Gleichgewicht zwischen Exploration und Exploitation zu verbessern, werden die Ausgleichskoeffizienten des KTGEL einer gleichmäßigen Randomisierung unterzogen. Der resultierende 'Random K-Tournament Grasshopper Extreme Learner (RKTGEL)' ist ein neuartiger Klassifikator mit einer gleichzeitig automatisierten Merkmalsselektion.
In einem ersten Schritt und basierend auf der konventionellen ELM-Methode wurden die Anzahl der Sensoren und deren Positionen untersucht. Für FMG hat sich gezeigt, dass 8 Sensoren geeignet sind, für sEMG sind nur 2 Sensoren geeignet und für sEIM sind 4 äquidistante Elektroden für Messungen im Frequenzbereich von 1 kHz bis 4 kHz geeignet. In einem zweiten Schritt wurden verschiedene Kollektionen von Handzeichen mit reduzierter Mehrdeutigkeit, mittlerer Mehrdeutigkeit und großer Mehrdeutigkeit definiert und durch die neuartigen Algorithmen einer Klassifizierung unterzogen. Kombinationen zwischen sEMG und FMG bzw. zwischen sEIM und FMG erreichen dabei in den Fällen der klein- und mittelmehrdeutigen Zeichensammlung selbst bei Daten von mindestens 20 Probanden eine Genauigkeit von insgesamt 100%. Für den Fall einer hohen Mehrdeutigkeit ist jedoch eine gezielte Reduktion der Mehrdeutigkeit durch Ausschluss von Handzeichen mit hoher Ähnlichkeit notwendig. Aus der Menge von 20 Gesten mit hoher Mehrdeutigkeit und nach Ausschluss von jeweils einem Handzeichen aus 6 mehrdeutigen Gestenpaaren und dem mehrdeutigen Zeichen ’Z’ blieben 13 Zeichen, darunter Buchstaben und Zahlen, die von 40 Probanden mit 2 sEMG- und 8 FMG-Sensoren erhoben wurden. Nach der Reduzierung der Mehrdeutigkeit erreichte das RKTGEL eine Gesamtgenauigkeit von 97%.:1 Introduction
2 Hand Gesture Recognition based on Myography
3 Extreme Learning Machine
4 Random K-Tournament Grasshopper Extreme Learner
5 Data Collection and Pre-Processing
6 Implementation of the Random K-Tournament Grasshopper Extreme Learner
7 Summary and Outloo
Embedded System Optimization of Radar Post-processing in an ARM CPU Core
Get PDFAlgorithms executed on the radar processor system contributes to a significant performance bottleneck of the overall radar system. One key performance concern is
the latency in target detection when dealing with hard deadline systems. Research has shown software optimization as one major contributor to radar system performance
improvements. This thesis aims at software optimizations using a manual and automatic approach and analyzing the results to make informed future decisions
while working with an ARM processor system. In order to ascertain an optimized implementation, a question put forward was whether the algorithms on the ARM
processor could work with a 6-antenna implementation without a decline in the performance. However, an answer would also help project how many additional
algorithms can still be added without performance decline.
The manual optimization was done based on the quantitative analysis of the software execution time. The manual optimization approach looked at the vectorization
strategy using the NEON vector register on the ARM CPU to reimplement the initial Constant False Alarm Rate(CFAR) Detection algorithm. An additional
optimization approach was eliminating redundant loops while going through the Range Gates and Doppler filters. In order to determine the best compiler for automatic
code optimization for the radar algorithms on the ARM processor, the GCC and Clang compilers were used to compile the initial algorithms and the optimized
implementation on the radar post-processing stage.
Analysis of the optimization results showed that it is possible to run the radar post-processing algorithms on the ARM processor at the 6-antenna implementation
without system load stress. In addition, the results show an excellent headroom margin based on the defined scenario. The result analysis further revealed that the
effect of dynamic memory allocation could not be underrated in situations where performance is a significant concern. Additional statements from the result demonstrated
that the GCC and Clang compiler has their strength and weaknesses when used in the compilation. One limiting factor to note on the optimization using the
NEON register is the sample size’s effect on the optimization implementation. Although it fits into the test samples used based on the defined scenario, there might
be varying results in varying window cell size situations that might not necessarily improve the time constraints
Polyacrylonitrile-based Hierarchical Porous Carbons for Supercapacitors
No full textThe globally increasing energy demand that results from the rapid development of modern society has created intensive attention towards the importance of energy efficiency. The areas of energy storage and energy conversion have become one of the most important topics in scientific community at present. As new generation energy storage elements, supercapacitors have exhibited promising practical prospects in the information, transportation, electronics and other sectors due to their charge and discharge performance at high rate, high power density as well as long cycle life. Energy density, including gravimetric energy density, areal energy density and volumetric energy density, is one of the most critical indicator evaluating the performance of supercapacitors. The electrochemical performance of supercapacitors depends mainly on the electrochemical activities and kinetic properties of electrode materials. Carbonaceous materials are deemed to be highly promising, and therefore are extensively investigated energy storage materials for supercapacitors because of their environmental friendliness, low-cost production and outstanding chemical inertness during charging-discharging processes. The specific surface area has been long thought to be the main factor influencing the capacitance of carbonaceous materials. However, the pore structure is of similar importance. High specific surface areas are always arising from a high content of micropores. However, pore radii in the sub-nanometer range impede the ionic charge transfer ability significantly and thus cause a damping of capacitance.
In this thesis, hierarchical porous carbons and their composite materials were fabricated by using polyacrylonitrile as carbon precursor for a tailored step-by-step pore forming method, including phase inversion, CaCO3 activation and KOH activation. The materials were thoroughly characterized by XRD, SEM, TEM, BET, XPS and Raman spectroscopy to ascertain the chemical and structural features. The electrochemical properties were studied by cyclic voltammetry (CV), galvanostatic charge-discharge (GCD) and electrochemical impedance spectroscopy (EIS) in detail to analyze the pore effect, which strongly influence their electrochemical properties.
Porous carbons with high specific surface areas up to 2315 m2 g-1 and high pore volume of 1.9 cm3·g-1 were prepared. A step-wise pore forming method was employed to ensure a high specific surface area and high content of macro/mesopore at the same time. The relationship between pore structure, electrochemical capacitance and rate capability was investigated by changing the content of micropores. For a same specific surface area, a higher micropore content led to a lower capacitance and poorer rate capability. Based on these results, the capacitance was optimized to be 286.8 F g-1.
The areal energy density of the supercapacitors can be improved by increasing the mass loading in a certain area directly. However, insufficient electrochemical reaction may be caused by a lack of unhindered electrical and ionic charge transfer routes, resulting in inefficient material utilization. This problem is addressed by designing hierarchical pore structures with embedded conductive additives. Thus, hierarchical porous carbons were modified by embedding carbon nanotubes (CNTs), followed by coverage with thin layers of birnessite. Owing to the hierarchical pore design and the very high pore volume, the birnessite coverage did not cause pore blocking. At the same time, an intimate contact between carbon and birnessite was established. A high area energy density of 627.8 μWh·cm-2 was obtained based on an optimized mass loading of 13.9 mg cm-2.
The volumetric energy density of supercapacitors was determined by the density and porosity of active materials. Similarly, the dense active materials not always generate high specific capacitance because of an increased dead mass. However, too porous active materials do not provide sufficient volumetric capacitance due to a waste of space. Thus, density and porosity must be balanced by hierarchical pore structure design so that all pores are interconnected and can be accessed by ions. At the same time, the content of these pores should be as low as possible to save space. Based on the results, highly hierarchical porous carbons were synthesized and embedded into conductive carbon foam to combine electronic conductivity with ionic transfer. In that way, a volumetric energy density as high as 19.44 µWh cm-3 at a volumetric power density of 500 mW cm-3 was generated
Die Bewertung der Nachlässe. Erfahrungen aus dem Institut für Geschichte und Archiv der Karls-Universität in Prag
Get PDFDer Artikel fasst kurz zusammen, wie die Nachlässe von Wissenschaftlern in den tschechischen Archiven behandelt werden. Wir versuchen, die Fragen der Bewertung zu skizzieren. Dabei nutzen wir vor allem die Erfahrungen aus der Übernahme und Bearbeitung von Nachlässen im Institut für die Geschichte und Archiv der Karlsuniversität, wo mehr als 100 Nachlässe aufbewahrt ist.The article summarizes briefly, how the personal papers of scholars and scientists are treated in Czech archives, expecially in the question of selection of documents. The authors tried to use their experiences from the practice in the Institute of History and Archives of Charles, whose archivists take care about personal papers of more than 100 scholars and scientists
Poisson Induced Bending Actuator for Soft Robotic Systems
No full textThis paper deals with a novel active bending soft body that employs metamaterials and combines soft behavior, integrated actuation, low complexity and a high density of producible forces and moments. The presented concept consists of a tube-like structure with tailored, unconventional material properties which enable the generation of a bending deformation and/or moment when circumferential stress and/or strain is induced. Circumferential actuation can be generated by a difference in pressure between the internal and external surface of the tube or, alternatively, by distributed expansion actuators that act radially or tangentially (e.g. shape memory wires). In addition to an analytical model, this paper also presents a design procedure and deals with the implementation of the proposed concept in a functional prototype and its experimental characterization
DRadEsel – Schulungsmaterial Interviews
Get PDFDas vorliegende Schulungsmaterial unterstützt die halbstandardisierte Interviewbefragung Radfahrender zu sicherheitskritischen Situationen mit Hilfe des DRadEsel-Interviewleitfadens
Interpretable Approximation of High-Dimensional Data based on the ANOVA Decomposition
No full textThe thesis is dedicated to the approximation of high-dimensional functions from scattered data nodes. Many methods in this area lack the property of interpretability in the context of explainable artificial intelligence. The idea is to address this shortcoming by proposing a new method that is intrinsically designed around interpretability. The multivariate analysis of variance (ANOVA) decomposition is the main tool to achieve this purpose. We study the connection between the ANOVA decomposition and orthonormal bases to obtain a powerful basis representation. Moreover, we focus on functions that are mostly explained by low-order interactions to circumvent the curse of dimensionality in its exponential form. Through the connection with grouped index sets, we can propose a least-squares approximation idea via iterative LSQR. Here, the proposed grouped transformations provide fast algorithms for multiplication with the appearing matrices. Through global sensitivity indices we are then able to analyze the approximation which can be used in improving it further. The method is also well-suited for the approximation of real data sets where the sparsity-of-effects principle ensures a low-dimensional structure. We demonstrate the applicability of the method in multiple numerical experiments with real and synthetic data.:1 Introduction
2 The Classical ANOVA Decomposition
3 Fast Multiplication with Grouped Transformations
4 High-Dimensional Explainable ANOVA Approximation
5 Numerical Experiments with Synthetic Data
6 Numerical Experiments with Real Data
7 Conclusion
BibliographyDie Arbeit widmet sich der Approximation von hoch-dimensionalen Funktionen aus verstreuten Datenpunkten. In diesem Bereich leiden vielen Methoden darunter, dass sie nicht interpretierbar sind, was insbesondere im Kontext von Explainable Artificial Intelligence von großer Wichtigkeit ist. Um dieses Problem zu adressieren, schlagen wir eine neue Methode vor, die um das Konzept von Interpretierbarkeit entwickelt ist. Unser wichtigstes Werkzeug dazu ist die Analysis of Variance (ANOVA) Zerlegung. Wir betrachten insbesondere die Verbindung der ANOVA Zerlegung zu orthonormalen Basen und erhalten eine wichtige Reihendarstellung. Zusätzlich fokussieren wir uns auf Funktionen, die hauptsächlich durch niedrig-dimensionale Variableninteraktionen erklärt werden. Dies hilft uns, den Fluch der Dimensionen in seiner exponentiellen Form zu überwinden. Über die Verbindung zu Grouped Index Sets schlagen wir dann eine kleinste Quadrate Approximation mit dem iterativen LSQR Algorithmus vor. Dabei liefern die vorgeschlagenen Grouped Transformations eine schnelle Multiplikation mit den entsprechenden Matrizen. Unter Zuhilfenahme von globalen Sensitvitätsindizes können wir die Approximation analysieren und weiter verbessern. Die Methode ist zudem gut dafür geeignet, reale Datensätze zu approximieren, wobei das sparsity-of-effects Prinzip sicherstellt, dass wir mit niedrigdimensionalen Strukturen arbeiten. Wir demonstrieren die Anwendbarkeit der Methode in verschiedenen numerischen Experimenten mit realen und synthetischen Daten.:1 Introduction
2 The Classical ANOVA Decomposition
3 Fast Multiplication with Grouped Transformations
4 High-Dimensional Explainable ANOVA Approximation
5 Numerical Experiments with Synthetic Data
6 Numerical Experiments with Real Data
7 Conclusion
Bibliograph