1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	Enhanced techniques in non-color based tracker for a vision system
Authors:	Deshinta Arrova Dewi (P65158)
Supervisor:	Elankovan Sundararajan, Assoc. Prof. Dr.
Keywords:	Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia Color graphics Color computer graphics Algorithms
Issue Date:	31-Jan-2019
Description:	Object tracking is a major area of concern in vision system and has been extensively used in numerous applications. The success of using color information in object tracking has been continuously reported with less attention given to several issues, such as the lengthy process to calibrate color, color fading due to outdoor exposure, and others. The need of such application that does not necessarily rely on color information has seen a hike due to the mentioned issues. Many previous researches have suggested techniques that do not utilize color information. However, various challenges are yet to overcome, for example, detect objects in low illuminations, data association of similar objects and acceleration of I/O bound program. This thesis suggests novel approaches that applicable to objects with and without predefined marker. A combination of contour detection and geometric moment is proposed for objects with predefined marker. Data association technique that exploit Kalman Filter with Euclidean Distance and Hungarian Algorithm is proposed to tackle detection in low illumination environment. This technique is able to achieve accuracy up to 90% under 50 lux. For objects without marker, a combined methods is proposed that includes Optical Flow with Lucas Kanade measurement. To enhance the ability in tracking, enhanced Shi Tomasi corner detection is applied to improve the quality of initial point of detection. An iteration in the body of algorithm is introduced to solve this issue. As result, good points are produced and processed by the enhanced Shi Tomasi. The experiments demonstrate that the combined technique is able to track fast moving objects like autonomous robots while playing a soccer game. The fast movement is tracked, although the camera is not in a perpendicular position. However, in term of speed, this combined technique is affected by I/O latency that makes program has to wait frames arrival upon processing them. For example, across 1231 frames arrival it takes 25.510 seconds processing time in which CPU time is 12.937 seconds. This implies most of computation times is pre-empted by I/O latency. To improve this condition, a parallelization using two threads is proposed. The idea is to separate I/O thread from processing thread using producer-consumer concept with a shared queue. This concept is exploited in a way a producer thread captures frames and put them in a queue. The consumer thread fetches frames from the queue for processing. This technique may speed up the overall processing time of 1231 frames up to 1.7 times faster with performance improvement 43.52% and efficiency 88.52%. This thesis also compare the number of frames that have been processed by the algorithm over the processing time. The result shows improvement from 32.268 frames processed in one second by sequential program becomes 59.850 frames processed in one second by parallel program. The achievement is possible because the main processing thread does not directly linked to the I/O tasks. The above results generated from the experiments with public datasets that are available online. The proposed algorithms were implemented using object-oriented programming written in Python 3.6 and OpenCV 3.2. The experiments further confirm that the novel approaches in this thesis offers benefit to the development and improvement of noncolor based tracker in a vision system.,Ph.D.
Pages:	200
Call Number:	T385.D469 2019 3 tesis
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.