University of Sussex
Browse
Hassan,_Waqas.pdf (7.55 MB)

Video analytics for security systems

Download (7.55 MB)
thesis
posted on 2023-06-08, 14:11 authored by Waqas Hassan
This study has been conducted to develop robust event detection and object tracking algorithms that can be implemented in real time video surveillance applications. The aim of the research has been to produce an automated video surveillance system that is able to detect and report potential security risks with minimum human intervention. Since the algorithms are designed to be implemented in real-life scenarios, they must be able to cope with strong illumination changes and occlusions. The thesis is divided into two major sections. The first section deals with event detection and edge based tracking while the second section describes colour measurement methods developed to track objects in crowded environments. The event detection methods presented in the thesis mainly focus on detection and tracking of objects that become stationary in the scene. Objects such as baggage left in public places or vehicles parked illegally can cause a serious security threat. A new pixel based classification technique has been developed to detect objects of this type in cluttered scenes. Once detected, edge based object descriptors are obtained and stored as templates for tracking purposes. The consistency of these descriptors is examined using an adaptive edge orientation based technique. Objects are tracked and alarm events are generated if the objects are found to be stationary in the scene after a certain period of time. To evaluate the full capabilities of the pixel based classification and adaptive edge orientation based tracking methods, the model is tested using several hours of real-life video surveillance scenarios recorded at different locations and time of day from our own and publically available databases (i-LIDS, PETS, MIT, ViSOR). The performance results demonstrate that the combination of pixel based classification and adaptive edge orientation based tracking gave over 95% success rate. The results obtained also yield better detection and tracking results when compared with the other available state of the art methods. In the second part of the thesis, colour based techniques are used to track objects in crowded video sequences in circumstances of severe occlusion. A novel Adaptive Sample Count Particle Filter (ASCPF) technique is presented that improves the performance of the standard Sample Importance Resampling Particle Filter by up to 80% in terms of computational cost. An appropriate particle range is obtained for each object and the concept of adaptive samples is introduced to keep the computational cost down. The objective is to keep the number of particles to a minimum and only to increase them up to the maximum, as and when required. Variable standard deviation values for state vector elements have been exploited to cope with heavy occlusion. The technique has been tested on different video surveillance scenarios with variable object motion, strong occlusion and change in object scale. Experimental results show that the proposed method not only tracks the object with comparable accuracy to existing particle filter techniques but is up to five times faster. Tracking objects in a multi camera environment is discussed in the final part of the thesis. The ASCPF technique is deployed within a multi-camera environment to track objects across different camera views. Such environments can pose difficult challenges such as changes in object scale and colour features as the objects move from one camera view to another. Variable standard deviation values of the ASCPF have been utilized in order to cope with sudden colour and scale changes. As the object moves from one scene to another, the number of particles, together with the spread value, is increased to a maximum to reduce any effects of scale and colour change. Promising results are obtained when the ASCPF technique is tested on live feeds from four different camera views. It was found that not only did the ASCPF method result in the successful tracking of the moving object across different views but also maintained the real time frame rate due to its reduced computational cost thus indicating that the method is a potential practical solution for multi camera tracking applications.

History

File Version

  • Published version

Pages

172.0

Department affiliated with

  • Informatics Theses

Qualification level

  • doctoral

Qualification name

  • phd

Language

  • eng

Institution

University of Sussex

Full text available

  • Yes

Legacy Posted Date

2013-01-17

Usage metrics

    University of Sussex (Theses)

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC