Tiefe Neuronale Netze und autonomes Fahren

David Mainzer -- Robert Bosch Car Multimedia GmbH
Der Vortrag gibt eine Einführung in das Gebiet der Tiefen Neuronalen Netze und erklärt an einem Problem aus der Bildverarbeitung (Klassifizierung von Objekten im Bild) die Funktionsweise des Verfahrens. Es wird auf Software Frameworks und Embedded Hardware für das Trainieren bzw. Ausführen der Neuronalen Netze eingegangen.
VDE-Vortragsreihe, Hannover, 2017
Cover of my Phd Thesis

New Geometric Algorithms and Data Structures for Collision Detection of Dynamically Deforming Objects

David Mainzer
This thesis presents a collision detection approach, which works entirely without an acceleration data structure and supports rigid and soft bodies. Furthermore, we can compute inter-object and intra-object collisions of rigid and deformable objects consisting of many tens of thousands of triangles in a few milliseconds. To realize this, a subdivision of the scene into parts using a fuzzy clustering approach is applied. Based on that all further steps for each cluster can be performed in parallel and if desired, distributed to different GPUs. Tests have been performed to judge the performance of our approach against other state-of-the-art collision detection algorithms. Additionally, we integrated our approach into Bullet, a commonly used physics engine, to evaluate our algorithm.
Universitätsbibliothek Clausthal, Clausthal, 2015
Collision Detection Based on Fuzzy Clustering for Deformable Objects on GPUs

Massively Parallel Batch Neural Gas for Bounding Volume Hierarchy Construction

René Weller, David Mainzer, Abhishek Srinivas, Matthias Teschner and Gabriel Zachmann
Ordinary bounding volume hierarchy (BVH) construction algorithms create BVHs that approximate the boundary of the objects. In this paper, we present a BVH construction that instead approximates the volume of the objects with successively finer levels. It is based on Batch Neural Gas (BNG), a clustering algorithm that is known from machine learning. Additionally, we present a novel massively parallel version of this BNG-based hierarchy construction that runs completely on the GPU. It reduces the theoretical complexity of the sequential algorithm from O(n log n) to O(log² n) and also our CUDA implementation outperforms the CPU version significantly in practice.
Workshop in Virtual Reality Interactions and Physical Simulation (Vriphys), Bremen, Germany, September 24 - 25, 2014.
Collision Detection Based on Fuzzy Scene Subdivision

Collision Detection Based on Fuzzy Scene Subdivision

David Mainzer and Gabriel Zachmann
We present a novel approach to perform collision detection queries between rigid and/or deformable models. Our method can handle arbitrary de- formations and even discontinuous ones. For this, we subdivide the whole scene with all objects into connected but totally independent parts by a fuzzy clustering algorithm. Following, for every part our algorithm performs a Principal Com- ponent Analyses to achieve the best sweep direction for the Sweep-Plane step, which reduces the number of false positives greatly. Our collision detection algo- rithm processes all computations without the need of a bounding volume hierar- chy or any other acceleration data structure. One great advantage of this is that our method can handle the broad phase as well as the narrow phase within one single framework. Our collision detection algorithm works directly on all prim- itives of the whole scene, which results in a simpler implementation and can be integrated much more easily by other applications. We can compute inter-object and intra-object collisions of rigid and deformable objects consisting of many tens of thousands of triangles in a few milliseconds on a modern computer. We have evaluated its performance by common benchmarks.
GPU Computing and Applications, Singapore, 9 Oct 2013, ISBN-13 978-981-287-133-6
Collision Detection Based on Fuzzy Clustering for Deformable Objects on GPUs

Poster: Collision Detection Based on Fuzzy Clustering for Deformable Objects on GPUs (CDFC)

David Mainzer and Gabriel Zachmann
We present a novel Collision Detection Based on Fuzzy Clustering for Deformable Objects on GPU’s (CDFC) technique to perform collision queries between rigid and/or deformable models. Our method can handle arbitrary deformations and even discontinuous ones. With our approach, we subdivide the scene into connected but totally independent parts by fuzzy clustering, and therefore, the algorithm is especially well-suited to GPU’s. Our collision detection algorithm processes all computations without the need of a bounding volume hierarchy or any other acceleration data structure. One great advantage of this is that our method can handle the broad phase as well as the narrow phase within one single framework. We can compute inter-object and intra-object collisions of rigid and deformable objects consisting of many tens of thousands of triangles in a few milliseconds on a modern computer. We have evaluated its performance by common benchmarks. In practice, our approach is faster than earlier CPU- and/or GPU-based approaches and as fast as state-of-the-art techniques but even more scalable.
International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision (WSCG) - POSTER Proceedings, Plzen, Czech Republic, June 24 - 27, 2013. ISBN 978-80-86943-76-3
AVILUS_BuchCover

Kollisionserkennung und natürliche Interaktion in virtuellen Umgebungen

David Mainzer, Rene Weller and Gabriel Zachmann
Dieser neuartige Ansatz verzichtet auf die Verwendung einer Beschleunigungsdatenstruktur. Die Kollisionsberechnung erfolgt direkt auf dem Dreiecksnetz, was Approximationsfehler vermeidet. Durch die Verwendung des Dreiecksnetzes ist eine Kollisionserkennung zwischen deformierbaren Objekten möglich. Aufgrund der fortschreitenden Entwicklung im Bereich der Many-Core Architekturen, ist das massiv parallele Bearbeiten von Daten von großer Bedeutung. Somit ist das Ziel des Ansatzes, die Parallelisierung der Kollisions-Pipeline. Hierfür werden die Berechnungen vollständig auf den CO-Prozessor, welcher viele Berechnungen parallel ausführen kann, wie z.B. die GPU, ausgelagert. Die Berechnung der Bounding-Box der Dreiecke aus dem Dreiecksnetz und deren anschließende Sortierung kann vollständig parallelisiert werden. Aus der Sortierung der Bounding-Boxen kann eine Überlapp-Liste erstellt werden, welche alle möglichen Dreiecks-Kandidaten für eine Kollision beinhaltet. Die verbleibenden Kandidaten werden auf Kollision getestet und das Ergebnis zurück an die Simulation gegeben.
Kollisionserkennung und natürliche Interaktion in virtuellen Umgebungen. In Virtuelle Techniken im industriellen Umfeld, chapter 3.2 and 3.4, pages 33–38 and 114–116, ISBN: 978-3-642-20635-1, Springer 2011.
benchmarking scenarios

A Benchmarking Suite for 6-DOF Real Time Collision Response Algorithms

Rene Weller, David Mainzer, Gabriel Zachmann, Mikel Sagardia, Thomas Hulin, Carsten Preusche
A benchmarking suite for rigid object collision detection and collision response schemes. The proposed benchmarking suite can evaluate both the performance as well as the quality of the collision response. The former is achieved by densely sampling the configuration space of a large number of highly detailed objects; the latter is achieved by a novel methodology that comprises a number of models for certain collision scenarios. With these models, we compare the force and torque signals both in direction and magnitude. Our device-independent approach allows objective predictions for physically-based simulations as well as 6-DOF haptic rendering scenarios. In the results, we show a comprehensive example application of our benchmarks comparing two quite different algorithms utilizing our proposed benchmarking suite. This proves empirically that our methodology can become a standard evaluation framework.
Proceedings of the 17th ACM Symposium on Virtual Reality Software and Technology 2010 (VRST' 2010), Hong Kong, November 2010.
book cover

Two Case Studies for Jazzyk BSM

Michael Köster, Peter Novák, David Mainzer and Bernd Fuhrmann
Recently, we introduced Behavioural State Machines (BSM), a novel programming framework for development of cognitive agents with Jazzyk, its associated programming language and interpreter. The Jazzyk BSM framework draws a strict distinction between knowledge representation and behavioural aspects of an agent program. Jazzyk BSM thus enables synergistic exploitation of heterogeneous knowledge representation technologies in a single agent, as well as offers a transparent way for embedding cognitive agents in various simulated or physical environments. This makes it a particularly suitable platform for development of simulated, as well as physically embodied cognitive agents, such as virtual agents, or non-player characters for computer games. In this paper we report on Jazzbot and Urbibot projects, two case-studies we developed using the Jazzyk BSM framework in simulated environments provided by a first person shooter computer game and a physical reality simulator for mobile robotics respectively. We describe the underlying technological infrastructure of the two agent applications and provide a brief account of experiences and lessons we learned during the development.
Agents for Games and Simulations, Springer-Verlag Berlin, Heidelberg ©2009, ISBN: 978-3-642-11197-6
 

Jazzbot: A non-monotonically reasoning bot in a simulated 3D environment

Peter Novák, David Mainzer, Michael Köster and Bernd Fuhrmann
In our previous research we designed Jazzyk, a modular programming language for development of cognitive agent systems. Jazzyk obeys two basic design principles: 1) it allows for an easy integration of heterogeneous knowledge representation technologies, and 2) draws a strict distinction between modeling agent's knowledge and reasoning vs. its behaviours. To further drive the development of Jazzyk, we implemented Jazzbot, a softbot embodied in a simulated 3D environment of a computer game Nexuiz. The core of Jazzbot's belief base is implemented as a logic program interpreted in the semantics of Answer Set Programming, thus exploiting the power of non-monotonic reasoning. It is complemented by a Ruby language module for representing the bot's topological knowledge about the environment. Jazzbot thus demonstrates the synergistic effect of using heterogeneous, in this case declarative and object-oriented, KR technologies in a single agent system.
architectur of jznexuiz module

Implementierung eines autonomen Agenten in einer simulierten 3D-Umgebung -- Interaktion mit der Umwelt (Diploma-Theses)

David Mainzer
In our previous research we designed Jazzyk, a modular programming language for development of cognitive agent systems. Jazzyk obeys two basic design principles: 1) it allows for an easy integration of heterogeneous knowledge representation technologies, and 2) draws a strict distinction between modeling agent's knowledge and reasoning vs. its behaviours. To further drive the development of Jazzyk, we implemented Jazzbot, a softbot embodied in a simulated 3D environment of a computer game Nexuiz. The core of Jazzbot's belief base is implemented as a logic program interpreted in the semantics of Answer Set Programming, thus exploiting the power of non-monotonic reasoning. It is complemented by a Ruby language module for representing the bot's topological knowledge about the environment. Jazzbot thus demonstrates the synergistic effect of using heterogeneous, in this case declarative and object-oriented, KR technologies in a single agent system.
April 2008, Supervisor: Dix, Jürgen.
Zurück

Finde mich auf ...