Deep learning for object detection in video streams
Promotor(s) : Louppe, Gilles
Date of defense : 26-Jun-2019/27-Jun-2019 • Permalink :
|Deep learning for object detection in video streams
|Translated title :
|[fr] Apprentissage approfondi pour la détection d'objets dans les flux vidéo
|Date of defense :
|Committee's member(s) :
|Van Droogenbroeck, Marc
|Number of pages :
|[fr] Deep Learning, Computer Vision, Sensor
|Engineering, computing & technology > Computer science
|Research unit :
|Université de Liège, Liège, Belgique
|Master : ingénieur civil électricien, à finalité spécialisée en "signal processing and intelligent robotics"
|Master thesis of the Faculté des Sciences appliquées
[fr] ”Artificial intelligence is no match for natural stupidity” is the famous quote of the
brilliant Albert Einstein. Born in the fifties, artificial intelligence was already a source
of admiration. It has undergone a spectacular evolution in recent years and continues to
develop following technological and computer advances. Nowadays, artificial intelligence
plays a fundamental role in research and is ubiquitous in our daily lives.
The detection of objects in a video stream with deep learning having become accessible
and efficient thanks to the more and more performant processor, this principle will be
adapted in order to be able to use it as a door opening sensor. This thesis was conducted
in a sensor company called BEA. It is a leading manufacturer of sensing solutions for
automatic doors systems. The company was founded in 1965 and its headquarters are
located in Li`ege, Belgium.
As artificial intelligence is making great strides in our technologies, the company has
decided to do some experimentations with this process. The detection and classification
of objects in a video stream through deep learning is the main subject of this work. The
use of it in real time is a big challenge because it was necessary to handle the use of a
neural network accelerator while ensuring that the model used was not too consequent.
Despite this constraint, it was still necessary that the precision and the accuracy were
sufficient for the application.
In order to be independent of the position of the camera and its orientation for future
applications, a ground projection algorithm has been implemented. Then, to improve the
feasibility of the process, a Kalman filter will be integrated into the detected objects and
a tracking will be assigned to it.
Subsequently, open decision-making and people counting were implemented and all
was adapted to the Raspberry Pi 3 B + embedded system.
Finally, various evaluation tests were carried out to demonstrate the system’s fidelity
and its promising potential.
Description: Master's Thesis Report | BEA sensors | Charles Lentz
Size: 9.16 MB
Format: Adobe PDF
Cite this master thesis
The University of Liège does not guarantee the scientific quality of these students' works or the accuracy of all the information they contain.