Recently the extremely accelerated technological advancement made computer vision applicable for more and more fields. Computer vision has become a part of our everyday lives. Mobile phones are able to detect faces and other objects in order to optimize the quality of taken pictures, cars recognize traffic signs, pedestrians, and other cars, automatic monitoring systems ensure the safety of public places and traffic, photo galleries organize our pictures based on their content. There are several tasks, which are unsolvable, or hard to solve without computer vision. One of them is object tracking, especially when we are not allowed to place special markers on objects to be tracked.
My work focuses on deep neural network-based, real-time, markerless object tracking. In my case, the main aim of real-time object tracking is to provide the accurate position of objects for a high-level monitoring system. This monitoring system is a central safety logic for a demonstrator application, which is able to intervene (direct intervention, alert, etc.) in the observed system to ensure its safety. This requires a tracker system with high accuracy, high throughput, and low end-to-end latency.
The case study of my work is the Model-based Demonstrator for Smart and Safe systems (MoDeS3 in short) system. It is about an autonomous model locomotives moving in a centrally managed system. The system ensures the safe traffic of the locomotives, it recognizes and avoids dangerous situations.
This thesis goes through the design and implementation process of a high-precision, real-time, deep learning-based object tracking system for MoDeS3, which is deployed to an embedded device. Accuracy and performance measurements validate the system in order to ensure that it can fulfill the requirements of the domain specific problem. This work is not only about a solution for a domain-specific tracking problem; it also provides a methodology for solving similar tasks.