Spatial and temporal background modelling of non-stationary visual scenes
Metadata
Show full item recordAbstract
The prevalence of electronic imaging systems in everyday life has become increasingly apparent
in recent years. Applications are to be found in medical scanning, automated manufacture, and
perhaps most significantly, surveillance. Metropolitan areas, shopping malls, and road traffic
management all employ and benefit from an unprecedented quantity of video cameras for monitoring
purposes. But the high cost and limited effectiveness of employing humans as the final
link in the monitoring chain has driven scientists to seek solutions based on machine vision techniques.
Whilst the field of machine vision has enjoyed consistent rapid development in the last
20 years, some of the most fundamental issues still remain to be solved in a satisfactory manner.
Central to a great many vision applications is the concept of segmentation, and in particular,
most practical systems perform background subtraction as one of the first stages of video
processing. This involves separation of ‘interesting foreground’ from the less informative but
persistent background. But the definition of what is ‘interesting’ is somewhat subjective, and
liable to be application specific. Furthermore, the background may be interpreted as including
the visual appearance of normal activity of any agents present in the scene, human or otherwise.
Thus a background model might be called upon to absorb lighting changes, moving trees and
foliage, or normal traffic flow and pedestrian activity, in order to effect what might be termed in
‘biologically-inspired’ vision as pre-attentive selection. This challenge is one of the Holy Grails
of the computer vision field, and consequently the subject has received considerable attention.
This thesis sets out to address some of the limitations of contemporary methods of background
segmentation by investigating methods of inducing local mutual support amongst pixels
in three starkly contrasting paradigms: (1) locality in the spatial domain, (2) locality in the shortterm
time domain, and (3) locality in the domain of cyclic repetition frequency.
Conventional per pixel models, such as those based on Gaussian Mixture Models, offer no
spatial support between adjacent pixels at all. At the other extreme, eigenspace models impose
a structure in which every image pixel bears the same relation to every other pixel. But Markov
Random Fields permit definition of arbitrary local cliques by construction of a suitable graph, and
3
are used here to facilitate a novel structure capable of exploiting probabilistic local cooccurrence
of adjacent Local Binary Patterns. The result is a method exhibiting strong sensitivity to multiple
learned local pattern hypotheses, whilst relying solely on monochrome image data.
Many background models enforce temporal consistency constraints on a pixel in attempt to
confirm background membership before being accepted as part of the model, and typically some
control over this process is exercised by a learning rate parameter. But in busy scenes, a true
background pixel may be visible for a relatively small fraction of the time and in a temporally
fragmented fashion, thus hindering such background acquisition. However, support in terms of
temporal locality may still be achieved by using Combinatorial Optimization to derive shortterm
background estimates which induce a similar consistency, but are considerably more robust
to disturbance. A novel technique is presented here in which the short-term estimates act as
‘pre-filtered’ data from which a far more compact eigen-background may be constructed.
Many scenes entail elements exhibiting repetitive periodic behaviour. Some road junctions
employing traffic signals are among these, yet little is to be found amongst the literature regarding
the explicit modelling of such periodic processes in a scene. Previous work focussing on gait
recognition has demonstrated approaches based on recurrence of self-similarity by which local
periodicity may be identified. The present work harnesses and extends this method in order
to characterize scenes displaying multiple distinct periodicities by building a spatio-temporal
model. The model may then be used to highlight abnormality in scene activity. Furthermore, a
Phase Locked Loop technique with a novel phase detector is detailed, enabling such a model to
maintain correct synchronization with scene activity in spite of noise and drift of periodicity.
This thesis contends that these three approaches are all manifestations of the same broad
underlying concept: local support in each of the space, time and frequency domains, and furthermore,
that the support can be harnessed practically, as will be demonstrated experimentally.
Authors
Russell, David MarkCollections
- Theses [3706]