University of Twente Student Theses


Computationally Efficient Vision-based Robot Control

Terrivel, M (2017) Computationally Efficient Vision-based Robot Control.

[img] PDF
Abstract:Video target tracking systems is a trending research topic, with a plethora of applications emerging from recent studies, with both visual object detection and tracking disciplines being the most notable, principally on embedded platforms. Not only are they employed in various fields, but also remarkably combines several branches of studies, such as control engineering, video processing, and more recently, machine learning and sensor fusion. Autonomous vehicles are a notable example, being equipped with a variety of sensors, including cameras, and widely apply image processing and sensor fusion techniques, thus providing more concise and high-level information, which increases robustness and more importantly, certainty, on decision making. However, such techniques must respect real-time constraints, especially in terms of timing, due the fact a delay might have a high cost under certain circumstances. Currently, there is a broad interest in processing images with neural networks, which are superior in terms of performance and robustness in comparison to traditional image processing algorithms. Although the rapid development of image sensors in combination with neural network technology, computational power of the underlying platform is still a bottleneck, especially for embedded applications. Moreover, the platform is commonly responsible for multiple tasks, which might include a user interface, data processing, (digital) filtering and high-level control, thus cannot be fully dedicated to the neural network itself. Finally, modern system-on-chips comprise hardware accelerators and multiple processing cores, which enable embedded systems to accomplish the desired throughputs, and achieve better results and efficiency in comparison to pure software implementations. Most of the time, these SoCs are completely customizable and interaction between software and hardware is facilitated. In this thesis, the focus is on both implementation and evaluation of a computational efficient robot control, based on neural networks to detect and localize a specific target (another robot), on an embedded platform. Sensor fusion and Kalman filtering are addressed, with the latter being used to post-process the output of the neural network, meanwhile the former combines encoder, accelerometer and gyroscope data which is used to derive the ego-motion information of the robot. Moreover, a specific approach for estimating the ego-motion impact in terms of pixels is proposed and discussed. The neural network, however, is not the focus, thus is only briefly discussed. During development, computational complexity, project extensibility and parallel tasks were the main concern, with the former being reduced whenever possible.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Embedded Systems MSc (60331)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page