000 Allgemeines, Wissenschaft
Refine
Year of publication
- 2016 (2)
Document Type
- Master's Thesis (1)
- Researchpaper (1)
Language
- English (2)
Has Fulltext
- yes (2)
Is part of the Bibliography
- no (2)
Keywords
- Convolutional Neural Network (1)
- Deep Learning (1)
- Distributed Systems (1)
- Fully Convolutional Network (1)
- Logging (1)
- Monitoring (1)
- RGB-D (1)
- Scalability (1)
- Semantic Segmentation (1)
- Verteilte Systeme (1)
Institute
- FB 1: Druck und Medien (2) (remove)
In recent years new trends such as industry 4.0 boosted the research and
development in the field of autonomous systems and robotics. Robots collaborate and
even take over complete tasks of humans. But the high degree of automation requires
high reliability even in complex and changing environments. Those challenging
conditions make it hard to rely on static models of the real world. In addition to
adaptable maps, mobile robots require a local and current understanding of the scene.
The Bosch Start-Up Company is developing robots for intra-logistic systems, which
could highly benefit from such a detailed scene understanding. The aim of this work
is to research and develop such a system for warehouse environments. While the
possible field of application is in general very broad, this work will focus on the
detection and localization of warehouse specific objects such as palettes.
In order to provide a meaningful perception of the surrounding a RGB-D camera is
used. A pre-trained convolutional network extracts scene understanding in the form
of pixelwise class labels. As this convolutional network is the core of the application,
this work focuses on different network set-ups and learning strategies. One difficulty
was the lack of annotated training data. Since the creation of densely labeled images
is a very time consuming process it was important to elaborate on good alternatives.
One interesting finding was that it’s possible to transfer learning to a high extent from
similar models pre-trained on thousands of RGB-images. This is done by selective
interventions on the net parameters. By ensuring a good initialization it’s possible
to train towards a well performing model within few iterations. In this way it’s
possible to train even branched nets at once. This can also be achieved by including
certain normalization steps. Another important aspect was to find a suitable way
to incorporate depth-information. How to fuse depth into the existing model? By
providing the height over ground as an additional feature the segmentation accuracy
was further improved while keeping the extra computational costs low.
Finally the segmentation maps are refined by a conditional random field. The joint
training of both parts results in accurate object segmentations comparable to recently
published state-of-the-art models.
Talking about highly scalable and reliable sys-
tems, issues like logging and monitoring are often
disregarded. However, being able to manage to-
day’s software systems absolutely requires deep
knowledge about the current state of applications
as well as the underlying infrastructure. Extract-
ing and preparing debug information as well as
various metrics in a fast and clearly arranged
manner is an essential precondition in order to
handle this task.
Since we at Bertsch Innovation GmbH also
face increasing requirements concerning Media-
Cockpit as one of our core products, we decided
to establish a centralized logging infrastructure
in order to come up to the application’s evolution
towards a more and more distributed system.
In this paper, I want to describe the steps
that I have taken in order to setup a functioning
logging tool stack consisting of Elasticsearch,
Logstash and Kibana (usually abbreviated as ELK stack ). Besides outlining proper
setup and configuration, I will also discuss possi-
ble pitfalls as well as custom adjustments made
when ELK did not meet our demands.