Back to search

Democritus University of Thrace (DUTH)

Συστήματα ρομποτικής όρασης για εντοπισμό και παρακολούθηση στόχου

Abstract

dc:description

The robotic vision systems are a research field, which has flourished the last decades concerning the innovations and applications. Besides, robotic vision is connected with the development of the computers, which evolve extremely fast. Moreover, the vision systems are highly related to the image processing. The development of the cameras is a common practice nowadays and they can be found in video-devices, cell phones, digital cameras and web-cameras. The robot vision is the intersection of several other ones, including computer vision, image processing, robotics etc. A first definition of the robot vision could be the use of images in robotics. One camera and a robotic arm, with at least one degree of freedom, constitute one robotic vision system in its simplest form. The use of one single camera for extracting the environmental information does not constitute a robotic vision system. Many vision systems have been developed. The use of computer vision in robotics, such as pose estimation, vergence angle, image stabilization etc. and its fusion in a robotic system is the main scope and purpose of this thesis. Here we describe some applications for robot vision systems, in which two tools are utilized. These are the Kalman filter and the fuzzy logic. The filter’s recursiveness, makes it ideal not only for smoothing the data, but also for the elimination of the noise levels. Fuzzy logic is actually the interface between human beings and machines, where humans can provide commands through natural language representations, coded in simple ’if...then’ rules. The examined and presented methodologies are basic in the majority of robotic systems, including, image stabilization, vergence angle, pose estimation and visuo-inertial fusion. The methodologies are connected with the tools as depicted in the diagrams below. The present thesis is divided into three parts, the robot vision technologies, the tracking robotic vision systems and finally to the identifying robotic vision systems. The first chapter of the first part involves the mathematical formulations of the 3D projections onto the image plane, and some basic tasks are described, such as the camera calibration which is a crucial procedure in the vision systems. Furthermore, in the same part are referred some basic procedures for the image processing modules, such as the noise, the filters and the color models. Consequently, the optic flow techniques are described, in which the projected 3D motion vectors are calculated. In the last chapter of the first part the log-polar transformation is analyzed. The Cartesian images are transformed into polar ones and subsequently into log-polar ones, which maintain the same data with less volume. All of these technologies are utilized into the aforementioned methodologies, improving the processing time and the accuracy of the respective systems. The second part of the thesis is devoted into two tracking robot vision applications. These have been developed so as to increase both the qualitative and quantitative features of the respective procedures. The first one is the image stabilization. The image sequences and the video signals contain two movements, one desired and one not. The desired motion aims to constantly track a target preserving the images as stable as possible. The undesired one, which is present in almost any such signal, results into low quality video. This kind of motion, depending on the application, is caused by the jitters produced by several factors, such as a rough terrain. The image stabilization computes the undesired motions and compensates them without thought affecting the desired ones. Our contribution focuses into two characteristics of this methodology. The first one deals with the reduction of the computational times during the image processing, which is a resource demanding and a highly time consuming process. The second part concerns the optimal handling of this information so as to maximize the compensation of the fluctuations, while preserving the visual quality. To this end, the Cartesian images are transformed into log-polar ones and the optical flow is not applied to the whole frame but instead in four background image regions. These subimages were chosen based on the fact that the target of interest lies in the foreground and specifically in the image center. This transformation enables the images to maintain all the necessary data, while reducing their volume. A fuzzy Kalman filter handles the extracted information. The fuzzy rules interact with the recursive filter providing compensation vectors which cut off the undesired motion effects but still preserving the intended ones with high visual quality. The second tracking applications deal with a common problem in binocular systems. The vergence angle is the fixation of both cameras to the same target, in order to obtain s single only fused image. In opposite cases, the two images are merged, creating a confusing image as the scenery’s characteristics are not distinguishable. This procedure is repeated constantly as the human vision system does. The brain fuses the image pairs so as a single image to be interpreted. The biomimetic systems adopt the human procedure with cameras (eyes), camera sensors (retina), servos (eye muscles) and computer units (human brain). The main criteria for verge or diverge are the distance from the eyes and the degree of similarity between the two images. The similarity task has widely been developed, although none of these technique mimics the human processes. In our research, we use fuzzy logic. The two inputs to the fuzzy system is the depth of the scene, i.e. the distance from the eyes, and the image-pair similarity. The proposed system is computationally inexpensive, as there is no need for prior camera calibration. The similarity criterion is robust against light variations and it is improved when it interacts with the depth input. The depth is computed by the system’s encoders readings and is being calculated algebraically. The simple mathematical formulations, for the computation of both inputs, render the system capable for real-time operation. The third part of the thesis, deals with the identification of one target and the accurate estimate of its pose. The pose estimation is one of the most common and demanding tasks in robotics. The target’s pose, requires image processing to find the key features, from which the pose vector is computed via projection matrices and homographies. The more fine the pixel accuracy of these features is, the more precise the final pose vector estimation is. Our contribution to this field is based on the fact that the target is found from a known arrangement of colored markers. The proposed arrangement can take any shape and its cost is very low. The markers are identified from their color information and their motion vectors. Moreover, we propose a new method for color identification by utilizing many color spaces which enable the system to respond accurately even under extreme lighting and environmental conditions. Furthermore, a new method for handling occlusions is presented, which were leading the vision systems to failure. The adopted architecture has some key novel aspects, as the camera is mounted outside the system, providing flexibility to be applied in many systems and the target is observed better. The vision system is connected with an inertial measurement unit, which is embodied on the target. The data fusion is achieved via a Kalman filter, which exploits the complementary attributes of the two sensors. The filter has higher accuracy and robustness than each dependent senor, and for this reason it is being used as a benchmark for other fusion techniques. In the last chapter of our research, a new visuo-inertial fusion method is presented. The inertial measurement units have short-term accuracy while in long-term applications the introduced error is incremental. These sensors are too sensitive and, as a result, they still measure accelerations even when they are immobilized. The isolation and elimination of this drift error is challenging and much effort has been put to either enhance or optimize existent filters, such as the Kalman filter and its several formulations (Extended, Unscented, Iterative etc.). Our contribution to this field concerns the adoption of a three level hierarchical fuzzy system for visuo-inertial fusion. The first level of the proposed system aims to reduce the drift error from the acceleration data. In the second layer, the respective fuzzy logic unit receives as inputs the filtered data from the first level and the visual measurements. The fused measurements are then imported to the last stage along with the system’s previous state in order to provide recursiveness to the system. The hierarchical fuzzy system is capable of isolating the noise and providing accurate measurements even in the cases where one of the sensors suffers from extreme disturbances and noise. The system is being examined for all of its Membership function parameters, such as their type, shape, distribution and the different arrangement of its inputs. The optimal parameters are selected during this system analysis, based on the minimization for error compared to ground-truth measurements. The proposed application broaches the use of fuzzy logic and the hierarchical fuzzy systems for fusion tasks, especially to robotic vision systems which cooperate with inertial sensors. To sum up, in this thesis, two applications of vision systems for target tracking and two for target identification are presented. The common features of the applications are that they utilize traditional methodologies combined with tools for decision making, control and data fusion.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2010

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kyriakoulis, Nikolaos
  • Κυριακούλης, Νικόλαος

Subjects

dc:subject × 14

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/20664

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Kyriakoulis, Nikolaos; Κυριακούλης, Νικόλαος. Συστήματα ρομποτικής όρασης για εντοπισμό και παρακολούθηση στόχου. Democritus University of Thrace (DUTH), 2010. http://hdl.handle.net/10442/hedi/20664