Bayesian data fusion approaches for vehicle video analytics

Al-Hmouz, Rami

Bayesian data fusion approaches for vehicle video analytics

Al-Hmouz, Rami

Permalink

Publication Type:: Thesis
Issue Date:: 2007

Closed Access

	Filename	Description	Size
	01Front.pdf	contents and abstract	788.63 kB	Adobe PDF	View/Open
	02Whole.pdf	thesis	61.57 MB	Adobe PDF	View/Open

Copyright Clearance Process

Recently Added
In Progress
Closed Access

This item is closed access and not available.

Full metadata record

Field	Value	Language
dc.contributor.author	Al-Hmouz, Rami
dc.date.accessioned	2016-06-05T23:22:09Z
dc.date.available	2016-06-05T23:22:09Z
dc.date.issued	2007
dc.identifier.uri	http://hdl.handle.net/10453/44003
dc.description	University of Technology, Sydney. Faculty of Engineering.	en_AU
dc.description	NO FULL TEXT AVAILABLE. Access is restricted indefinitely. The hardcopy may be available for consultation at the UTS Library.
dc.description.abstract	NO FULL TEXT AVAILABLE. Access is restricted indefinitely. ----- Vehicle Video Analytics (WA) deals with the analysis of the data from a video associated with car images and automatically picking out the information that identifies the vehicle. For example, Licence Plate Recognition (LPR) is a one of the WA applications dealing with the identification of the plates automatically, however, it may not be enough to completely characterise the vehicle; vehicles' features include colour, model, size, speed, shape and number of axles. Combining multiple vehicles' features in a framework will be useful in many applications including stolen car retrieval, petrol theft, car park management and road asset protection. Inspired by these applications, the thesis introduces a comprehensive framework for a vehicle video analytics to identify the many features of vehicles. Several new and improved algorithms are proposed to extract and recognise vehicles features along with image enhancement methods to boost the over all performance of the WA systems. Most LPR systems perform well in a controlled environment or when the effect of noise on the image processing algorithms is small. However, in realistic situations, random noise occurs for many reasons; such as variations in lighting conditions, image quality and additional noise from image sensors. The impact of the random noise can be reduced using information from multiple frames/thresholds. This thesis proposes novel multi frame/threshold data fusion approaches to reduce the effect of the external and internal random noise. Licence plate localisation algorithms are exploited from thresholding techniques (Otsu, Hysteresis, etc.) in order to extract the plate from the image; however, the thresholding does not segment the plate under various illumination conditions. We propose an approach of threshold variation which ensures the object is to be segmented at least once, the dusters are modelled as a joint distribution of their features, and all data from different thresholds update the posterior using Bayesian inference. In the recognition part, we propose a novel approach for increasing the accuracy for any particular character recognition technique by taking advantage of the available frames. Each frame result acts as an individual sensor which informs its recognition results. These results from each individual frame (sensor) are fused to obtain the final recognition result. Simple approaches for the speed, model and colour of the vehicle based on the plate location are presented. The speed is estimated through out tacking the plate across frames. The approaches of plate extraction and recognition are exploited for emblem recognition as well. Since colours can be seen differently under different illuminations, we use illumination estimation and correction for the image before fusing the colour histograms of three areas around the plate which represent vehicle's colour. Extensive experimental test were carried using real data collected on high ways and car parks. Through out the thesis, the fusion problem formulation in the context of WA is illustrated and performance improvements are demonstrated.	en_AU
dc.format	Thesis (PhD)
dc.language.iso	en_AU	en_AU
dc.rights	The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
dc.rights	info:eu-repo/semantics/closedAccess
dc.title	Bayesian data fusion approaches for vehicle video analytics	en_AU
dc.type	Thesis	en_AU
utslib.copyright.status	closed_access

Abstract:

NO FULL TEXT AVAILABLE. Access is restricted indefinitely. ----- Vehicle Video Analytics (WA) deals with the analysis of the data from a video associated with car images and automatically picking out the information that identifies the vehicle. For example, Licence Plate Recognition (LPR) is a one of the WA applications dealing with the identification of the plates automatically, however, it may not be enough to completely characterise the vehicle; vehicles' features include colour, model, size, speed, shape and number of axles. Combining multiple vehicles' features in a framework will be useful in many applications including stolen car retrieval, petrol theft, car park management and road asset protection. Inspired by these applications, the thesis introduces a comprehensive framework for a vehicle video analytics to identify the many features of vehicles. Several new and improved algorithms are proposed to extract and recognise vehicles features along with image enhancement methods to boost the over all performance of the WA systems. Most LPR systems perform well in a controlled environment or when the effect of noise on the image processing algorithms is small. However, in realistic situations, random noise occurs for many reasons; such as variations in lighting conditions, image quality and additional noise from image sensors. The impact of the random noise can be reduced using information from multiple frames/thresholds. This thesis proposes novel multi frame/threshold data fusion approaches to reduce the effect of the external and internal random noise. Licence plate localisation algorithms are exploited from thresholding techniques (Otsu, Hysteresis, etc.) in order to extract the plate from the image; however, the thresholding does not segment the plate under various illumination conditions. We propose an approach of threshold variation which ensures the object is to be segmented at least once, the dusters are modelled as a joint distribution of their features, and all data from different thresholds update the posterior using Bayesian inference. In the recognition part, we propose a novel approach for increasing the accuracy for any particular character recognition technique by taking advantage of the available frames. Each frame result acts as an individual sensor which informs its recognition results. These results from each individual frame (sensor) are fused to obtain the final recognition result. Simple approaches for the speed, model and colour of the vehicle based on the plate location are presented. The speed is estimated through out tacking the plate across frames. The approaches of plate extraction and recognition are exploited for emblem recognition as well. Since colours can be seen differently under different illuminations, we use illumination estimation and correction for the image before fusing the colour histograms of three areas around the plate which represent vehicle's colour. Extensive experimental test were carried using real data collected on high ways and car parks. Through out the thesis, the fusion problem formulation in the context of WA is illustrated and performance improvements are demonstrated.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/44003