Refine
Year of publication
Document Type
- Conference Proceeding (9)
- Article (3)
- Doctoral Thesis (1)
Keywords
- 3D object detection (1)
- Anomaly detection (1)
- Building Automation (1)
- CCD-Bildwandler (1)
- CO2 (1)
- Carbon Dioxide (1)
- Datasets (1)
- Deep learning (1)
- Dreidimensionale Bildverarbeitung (1)
- Education (1)
- Klassifikator <Informatik> (1)
- LiDAR (1)
- Neural networks (1)
- Objekterkennung (1)
- Process optimization (1)
- Quality control (1)
- Smart Building (1)
- Three-dimensional displays (1)
- Vorverarbeitung (1)
- autonomous driving (1)
In this paper we investigate the use of deep neural networks for 3D object detection in uncommon, unstructured environments such as in an open-pit mine. While neural nets are frequently used for object detection in regular autonomous driving applications, more unusual driving scenarios aside street traffic pose additional challenges. For one, the collection of appropriate data sets to train the networks is an issue. For another, testing the performance of trained networks often requires tailored integration with the particular domain as well. While there exist different solutions for these problems in regular autonomous driving, there are only very few approaches that work for special domains just as well. We address both the challenges above in this work. First, we discuss two possible ways of acquiring data for training and evaluation. That is, we evaluate a semi-automated annotation of recorded LIDAR data and we examine synthetic data generation. Using these datasets we train and test different deep neural network for the task of object detection. Second, we propose a possible integration of a ROS2 detector module for an autonomous driving platform. Finally, we present the performance of three state-of-the-art deep neural networks in the domain of 3D object detection on a synthetic dataset and a smaller one containing a characteristic object from an open-pit mine.
In der Vergangenheit basierten große Systemintegrationsprojekte in der Regel auf Individualentwicklungen für einzelne Kunden. Getrieben durch Kostendruck steigt aber der Bedarf nach standardisierten Lösungen, die gleichzeitig die individuellen Anforderungen des jeweiligen Umfelds berücksichtigen. T-Systems GEI GmbH wird beiden Anforderungen mit Produktkerneln gerecht. Neben den technischen Aspekten der Kernelentwicklung spielen besonders organisatorische Aspekte eine Rolle, um Kernel effizient und qualitativ hochwertig zu entwickeln, ohne deren Funktionalitäten ins Uferlose wachsen zu lassen. Umgesetzt hat T-Systems dieses Konzept für Flughafeninformationssysteme. Damit kann dem wachsenden Bedarf der Flughafenbetreiber nach einer effizienten und kostengünstigen Softwarelösung zur Unterstützung Ihrer Geschäftsprozesse entsprochen werden.
Der Erfolg eines Softwarenentwicklungsprojektes insbesondere eines Systemintegrationsprojektes wird mit der Erfüllung des „Teufelsdreiecks“, „In-Time“, „In-Budget“, „In-Quality“ gemessen. Hierzu ist die Kenntnis der Software- und Prozessqualität essenziell, um die Einhaltung der Qualitätskriterien festzustellen, aber auch, um eine Vorhersage hinsichtlich Termin- und Budgettreue zu treffen. Zu diesem Zweck wurde in der T-Systems Systems Integration ein System aus verschiedenen Key Performance Indikatoren entworfen und in der Organisation implementiert, das genau das leistet und die Kriterien für CMMI Level 3 erfüllt.
Aim of the AXON2 project (Adaptive Expert System for Object Recogniton using Neuml Networks) is the development of an object recognition system (ORS) capable of recognizing isolated 3d objects from arbitrary views. Commonly, classification is based on a single feature extracted from the original image. Here we present an architecture adapted from the Mixtures of Eaqerts algorithm which uses multiple neuml networks to integmte different features. During tmining each neural network specializes in a subset of objects or object views appropriate to the properties of the corresponding feature space. In recognition mode the system dynamically chooses the most relevant features and combines them with maximum eficiency. The remaining less relevant features arz not computed and do therefore not decelerate the-recognition process. Thus, the algorithm is well suited for ml-time applications.
This paper addresses the pixel based classification of three dimensional objects from arbitrary views. To perform this task a coding strategy, inspired by the biological model of human vision, for pixel data is described. The coding strategy ensures that the input data is invariant against shift, scale and rotation of the object in the input domain. The image data is used as input to a class of self organizing neural networks, the Kohonen-maps or self-organizing feature maps (SOFM). To verify this approach two test sets have been generated: the first set, consisting of artificially generated images, is used to examine the classification properties of the SOFMs; the second test set examines the clustering capabilities of the SOFM when real world image data is applied to the network after it has been preprocessed to be invariant against shift, scale and rotation. It is shown that the clustering capability of the SOFM is strongly dependant on the invariance coding of the images.
This paper addresses the pixel based recognition of 3D objects with bidirectional associative memories. Computational power and memory requirements for this approach are identified and compared to the performance of current computer architectures by benchmarking different processors. It is shown, that the performance of special purpose hardware, like neurocomputers, is between one and two orders of magnitude higher than the performance of mainstream hardware. On the other hand, the calculation of small neural networks is performed more efficiently on mainstream processors. Based on these results a novel concept is developed, which is tailored for the efficient calculation of bidirectional associative memories. The computational efficiency is further enhanced by the application of algorithms and storage techniques which are matched to characteristics of the application at hand.
This paper describes the realization of a novel neurocomputer which is based on the concepts of a coprocessor. In contrast to existing neurocomputers the main interest was the realization of a scalable, flexible system, which is capable of computing neural networks of arbitrary topology and scale, with full independence of special hardware from the software's point of view. On the other hand, computational power should be added, whenever needed and flexibly adapted to the requirements of the application. Hardware independence is achieved by a run time system which is capable of using all available computing power, including multiple host CPUs and an arbitrary number of neural coprocessors autonomously. The realization of arbitrary neural topologies is provided through the implementation of the elementary operations which can be found in most neural topologies.
