AN AUTOMATIC APPROACH FOR PARAMETER SELECTION IN

SELF-ADAPTIVE TRACKING

Daniela Hall, R

emi Emonet and James L. Crowley

INRIA Rh

one-Alpes

France

Keywords:

Tracking, performance optimization, automatic parameter regulation.

Abstract:

In this article we propose an automatic approach for parameter selection of a tracking system. We show that

such a self-adaptive tracking system achieves better tracking performance than a system with manually tuned

parameters. Our approach requires little supervision by a user which makes this approach ideally suited for

commercial applications. The self-adaptive component makes the system less sensitive to changing environ-

mental conditions. Components for tracking, auto-critical evaluation and automatic parameter regulation serve

to detect performance drops that trigger the parameter regulation process. The self-adaptive components re-

quire a quality measure based on a statistical scene reference model. We propose an automatic approach for

the generation of such a reference model and compare several learning approaches. The experiments show

that the auto-regulation of parameters signiﬁcantly enhances the performance of the tracking system.

1 INTRODUCTION

In this article we propose an automatic approach for

parameter selection of a tracking system. We show

that such a self-adaptive tracking system achieves bet-

ter tracking performance than a system with manually

tuned parameters. Our approach requires little super-

vision by a user during initialisation which makes this

approach ideally suited for commercial applications.

The future goal of this work is to incorporate the

technology of self-adaptation into commercial video

surveillance systems. The growing number of in-

stalled video surveillance systems represent a great

demand for automatic setup and conﬁguration. Nowa-

days, an engineer is required for system installation.

The installation requires also the manual adaptation

of parameters. Such systems perform well as long

as the environment stays constant. Unfortunately, in

most real applications the environmental conditions

perceived by the sensors frequently change, which of-

ten breaks the system and requires reinitialisation and

new hand tuning of the parameters. For a company

with a large number of systems, the maintenance in-

cluding manual parameter tuning would require too

much resources. This article proposes an approach

that makes hand tuning of parameters obsolete. The

approach is automatic that means it requires no hu-

man supervision during run time, human supervision

is required only for the initialisation of the model gen-

eration.

Our approach generates a scene reference model

that models correct system output in a probabilistic

manner. Based on this model we deﬁne a metric to

judge the system output (detect errors and detect per-

formance decrease). This metric is also used to select

the optimal parameter setting for the current environ-

mental conditions. The selection is automatic with-

out prior knowledge about the environmental condi-

tion change.

In this article we address the important steps of our

approach. First, generation of the scene reference

model and model selection. Second, deﬁnition of a

quality metric and third, development of a strategy for

parameter space exploration.

As related work we want to cite Murino who ad-

dresses the problem of automatic parameter regu-

lation for vision systems in (Murino et al., 1996).

He proposes a multi-layered component architecture.

Each layer has its own set of parameters that are tuned

such that the evidence (coming from the lower level)

and the expectation (coming from the higher level) are

consistent. The improvements are not convincing and

the approach lacks the use of an external knowledge

base.

Hall D., Emonet R. and L. Crowley J. (2006).

AN AUTOMATIC APPROACH FOR PARAMETER SELECTION IN SELF-ADAPTIVE TRACKING.

In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 20-26

DOI: 10.5220/0001372600200026

 SciTePress

Scotti (Scotti et al., 2003) proposes an approach

based on self organizing maps (SOM). The SOM is

learnt by registering good parameter settings. During

run time, the automatic parameter selection chooses

the closest setting in SOM space that performed best

during training. The experiments are not convincing

and we think that the method has strong limitations.

In (Min et al., 2004), Min proposes an approach

for comparing the performance of different segmen-

tation algorithms by searching the optimal parame-

ters for each algorithm. He proposes an interesting

multi-loci hill climbing scheme on a coarsely sam-

pled parameter space. The segmentation system per-

formance is evaluated with respect to a given ground

truth. This approach is designed for the comparison

of algorithms and requires to test a large number of

different parameter settings. For this reason, the util-

ity of this approach for on-line parameter regulation

is less appropriate.

The remainder of the article is organised as fol-

lows. Section 2 describes the components of the self-

adaptive system architecture. The self-adaptive com-

ponents depend on a quality metric described in sec-

tion 3. Section 4 describes experiments in which we

test the self-adaptive capabilities of our system. We

ﬁnish with an overview and an outlook on how this

kind of system could be used in commercial applica-

tions.

2 ARCHITECTURE OF A

SELF-ADAPTIVE TRACKING

SYSTEM

In this section, we propose an architecture for a self-

adaptive tracking system. We ﬁrst discuss the archi-

tecture and then describe the individual components.

Commercial use of automatic parameter tuning has

very strong requirements on the method. First, the

method should reduce user interaction to a minimum

and second, parameter tuning should not slow down

tracking of the master system. This architecture sepa-

rates the master system from the self-adaptive compo-

nent which makes it suited for an implementation on a

distributed system. The tracking system with the fast

monitoring component can run on one host and the

costly parameter tuning process can run on a second

host while the master system continues tracking.

2.1 Self-adaptive Systems

Robertson and Brady (Robertson and Brady, 1999)

propose an architecture for self-adaptive systems.

They consider an image analysis system as a closed-

loop control system that integrates knowledge in or-

Auto−critical

evaluation

module

Regulation

reference model

output

score

Robust tracker

Self−adaptive tracking system

Post processing

Control Component

optimized parameters

regulation

request

output

Supervisor

Target

detection

Robust

tracking

Figure 1: Architecture of a self-adaptive tracking system.

der to be self-evaluating. Measuring and comparing

the system output to the desired output and applying

a corrective force to the system leads to increased per-

formance. The difﬁcult point is to generate a model of

the desired output. They demonstrate their approach

on the segmentation of aerial images using a bank of

different ﬁlter operators. The system selects automat-

ically the best ﬁlter for the current image conditions.

We follow this line of research and design a self-

adaptive system with a control component. The con-

trol component implements following abilities:

• Auto-critical evaluation: This means that the sys-

tem is able to judge its own performance.

• Auto-regulation of parameters: The ability to auto-

matically adapt the system’s parameters to the cur-

rent environmental conditions and ensure constant

performance.

2.2 System Architecture

Figure 1 shows the architecture of the self-adaptive

tracking system containing a tracking system and an

independent control component. The independence

of the control component allows to endow other types

of perceptual systems with self-adaptive capabilities

in a plug-and-play manner.

The above abilities of auto-critical evaluation and

auto-regulation are implemented as modules within

the control component. The output of the perceptual

system is monitored by the module for auto-critical

evaluation. Auto-regulation is triggered by an exter-

nal request from the post processing module. The

modules have access to a common knowledge base

that contains the reference model.

The structure and content of this knowledge base

is task dependent. This means that each system setup

requires a knowledge base with different structure and

different content. Section 3 explains in detail how this

knowledge base can be generated automatically in the

context of a robust tracking system.

2.3 Robust Real-time Tracking

This self-adaptive architecture is applied to a real-

time tracking system in a video surveillance scenario.

AN AUTOMATIC APPROACH FOR PARAMETER SELECTION IN SELF-ADAPTIVE TRACKING

The tracking system is composed of a central super-

visor that calls in a cycle the modules video demon,

automatic target detection and robust tracking. The

supervisor manages the list of current targets. The

tracking module provides robust tracking of the cur-

rent targets using a Kalman ﬁlter. The detection mod-

ule based on adaptive background differencing de-

tects new targets that are added to the target list. The

system can track up to 8 targets in images of 384×288

pixels at 30Hz on a 2 GHz processor.

The robust tracking system produces for each

frame t

and each target a vector (measurement) com-

posed of centroid and width and height of the target’s

bounding box y(t

)=(x

,w,h)

. These vec-

tors are summarised in a log ﬁle in XML format. The

tracking result depends on a number of parameters

such as detection energy threshold (minimum target

size), noise threshold (pixel energy below this thresh-

old is considered as noise) and parameters that control

split and merge of targets. These parameters deter-

mine how close targets need to be for fusion or sepa-

ration. For further details on the system implementa-

tion see (Caporossi et al., 2004).

2.4 Auto-critical Evaluation

The goal of the auto-critical evaluation is to monitor

the performance of the system and detect performance

drops. This requires the deﬁnition of a measure that

estimates the goodness of trajectories (measurement

sequences) with respect to a reference model that is

constructed in a learning phase. Such a reference

model describes what usually happens in the scene.

An example of a different reference model is the se-

mantic scene model of Makris and Ellis in (Makris

and Ellis, 2003) where they learn entry and exit points

from examples and represent them as a Gaussian mix-

ture. Trajectories are represented by a topological

graph.

There is a wide range of different representation

forms of reference models. In addition to graphs and

Gaussian Mixture Models (GMMs), we want to name

histograms that often provide a good solution to con-

crete problems despite their simplicity. Probability

density approximations such as histograms or GMMs

have the advantage that a goodness score can be de-

ﬁned easily based on statistical estimation.

All probabilistic reference models have in common

that they estimate the true probability density func-

tion (pdf) of measurements y =(x

,w,h)

.For

example a pdf represented by a GMM can be ob-

tained by applying a standard learning approach such

as clustering to the training data and representing each

cluster by a Gaussian. In that case, the probability

density of a d dimensional measurement y is com-

puted by

p(y)=



j=1

p(y|C

)P (C

) (1)

p(y|C

)=p(y| µ

)

(2π)

d/2

|U|

1/2

(−0.5(y− µ

)

−1

(y− µ

))

(2)

with µ

mean and U

covariance of Gaussian C

. The

priors P (C

) are estimated from the training data:

P (C

) ≈

(3)

with |C

| number of data points associated to C

dur-

ing training and M total number of data points used

for training.

Equation 1 provides the probability density of sin-

gle measurements. To compute the quality of a tra-

jectory, we average the probability of the single mea-

surements. The goodness G

avg

(y(t)) of the trajectory

y(t)=(y

,...,y

) with length n +1is computed as

follows:

avg

(y(t)) =

n +1



i=0

p(y

) (4)

using eq 1 for computing p(y

). We have tested two

other goodness measures (see (Hall, 2005)), but the

simple averaging technique provided the best results.

This monitoring component is able to detect errors

and global performance drops automatically by eval-

uating the goodness score of the system output.

2.5 Automatic Parameter Regulation

The goal of the regulation module is to ﬁnd a para-

meter setting that increases the system performance.

In the current implementation, this requires the sim-

ulation of tracking output using different parameter

settings. Depending on the number of tested parame-

ter settings, this task is time consuming. In a real-time

commercial application, the tracking application and

the parameter regulation can be run in parallel on a

distributed system.

The module architecture is illustrated in Figure 2.

Parameter regulation is started by a request from the

post processing module in cases where a performance

drop is detected. The request contains the current pa-

rameter setting and an image sequence composed of

the last k frames. The regulation tool searches now

for a parameter setting that has good performance on

this sample sequence. The parameter space explo-

ration tool provides new parameter settings. Feedback

of the goodness of previous parameter settings may be

used to guide the search. The tracking system simula-

tor simulates the output of the tracking system on the

sample sequence with these parameter settings.

VISAPP 2006 - IMAGE UNDERSTANDING

Parameter space

exploration tool

Robust tracker simulator

Regulation module

Regulation request with image sequence and current parameters

reference model

Auto−critical

evaluation

Supervisor

Target

detection

Robust

tracking

best

score

output

parameters

Image sequence

Figure 2: Architecture of the automatic parameter regula-

tion module.

The goodness score for the output is computed by

the auto-critical evaluation. Ideally, the parameter

regulation continues until it ﬁnds a parameter setting

that repairs the failure (this means that the perfor-

mance must exceed a predeﬁned acceptance thresh-

old). In the experiments, auto-regulation of the pa-

rameters is performed on the entire sequence. The

current implementation tests a maximum number of

parameter settings and returns the one with the best

performance.

The parameter regulation module contains an in-

dependent module for parameter space exploration.

This allows to test different exploration strategies. We

tested an enumerative strategy and a strategy based on

a genetic algorithm. Gradient based methods or adap-

tive sampling as in (Min et al., 2004) could also be

used.

3 GENERATION OF A SCENE

REFERENCE MODEL

The scene reference model together with a quality

metric forms the knowledge base of the self-adaptive

system. It allows the system to judge the quality of

the system output and to select parameters that are

optimal with respect to this metric. The success of

the self-adaptive technique depends on the represen-

tativeness of this scene reference model and its metric.

As a consequence, model generation is an important

step within this approach.

With respect to commercial applications and the

fact that only a limited number of ground truth data

may be available for initialisation, we focus espe-

cially on model representation forms and learning

techniques that are incremental. Such techniques have

the great advantage, that they can be reﬁned as more

data becomes available.

In this section, we explain different model gen-

eration methods. All methods operate on the same

4 dimensional training data of the form y =

,w,h)

and provide an estimation of the un-

derlying pdf.

3.1 Learning Methods for Model

Generation

We tested several incremental and non-incremental

methods for model generation from hand-labelled

training data. The most interesting methods for which

we show experimental results are:

• non incremental methods like kmeans with pruning

(KM-F) and EM with pruning (EM-F),

• incremental method based on a multi-dimensional

histogram Histo,

• the incremental method Histo-EM-F consists in

building a ﬁne grain histogram, extracting a

weighted point for each non empty cell and then

performing EM with pruning on the extracted

points. The histogram serves to obtain an interme-

diate representation of the data and to reduce the

number of points considered for EM.

The current implementation requires hand labelled

data for model construction. In a future version, we

envision an iterative learning scheme that generates

an initial model from little hand labelled data (a few

trajectories). This model would then be used to ﬁlter

the output of the tracking system to obtain more cor-

rect data. An improved model would then be learnt

from the enhanced set of training data. Several iter-

ations of this scheme should produce a high quality

model.

In such an iterative learning scheme, an incremen-

tal model structure has several advantages. Incremen-

tal algorithms have the advantage of being able to re-

ﬁne the model subsequently while new observations

arrive. Incremental models can incorporate very large

amounts of data, because training data from previous

iterations does not need to be stored. For this rea-

son, incremental learning schemes should be able to

produce higher quality models than non-incremental

learning schemes.

3.2 Reference Model Selection

The above methods have several parameters such as

the number of Gaussians in the GMM that can not be

determined a priori. A successful strategy for ﬁnd-

ing a model of high quality is to generate a large set

of models by varying the parameters of the methods

and then selecting the best model with respect to some

quality criterion. We propose to use a quality criterion

based on the probability of classiﬁcation error with re-

spect to a validation set of positive and negative mea-

surements. Section 4 describes how such a validation

set can be acquired.

AN AUTOMATIC APPROACH FOR PARAMETER SELECTION IN SELF-ADAPTIVE TRACKING

For each model, a classiﬁer is built by selecting a

threshold on the probability density p(y) of the mea-

surements y. Given the validations sets Posand Neg

of positive and negative examples, the threshold is

chosen automatically as the value that minimise the

probability of classiﬁcation error:

err

|Pos| + |Neg|

min

⎛

⎝



p∈Pos

pdf(p)<th



n∈Neg

pdf(n)≥th

⎞

⎠

(5)

where δ is the Kronecker symbol. The least complex

model with a P

err

value below an acceptance thresh-

old is said to have the best quality. This quality mea-

sure based on P

err

is an efﬁcient way of selecting the

model.

The complexity of a model is related to the num-

ber of Gaussians or the the number of histogram cells

(depending on the representation). Selecting the least

complex model that fulﬁls an acceptance threshold

is a regularisation method that avoids over ﬁtting.

The acceptance threshold is determined experimen-

tally (see Section 4).

4 EXPERIMENTAL EVALUATION

In the experiments we want to demonstrate three

points: i) the quality of the automatic parameter selec-

tion, ii) the performance increase of tracking by auto-

regulation of parameters compared to manual parame-

ter tuning and iii) the link between the model quality

measure P

err

and the system performance.

4.1 Performance Evaluation

For measuring the performance of our tracking sys-

tem, we measure recall and precision. These val-

ues are computed by a statistics tool that operates on

XML description of the bounding boxes of the de-

tected targets and of the manually annotated targets.

The statistics tool determines for each frame the best

matching pairs with respect to overlap of detected and

ground truth boxes. For a particular overlap thresh-

old T , the tool evaluates the number of true positives

(TP), false positives (FP) and false negatives (FN).

Recall =

TP + FN

Precision =

TP + FP

(6)

The overlap of two bounding boxes is computed by:

O(A

obs

truth

λ {A

obs

∩ A

truth

}

λ {A

obs

∪ A

truth

}

(7)

where λ {...} represents the surface (Lebesgue mea-

sure in dimension 2) and A the bounding box.

Figure 3: Evaluation of the quality of automatically selected

parameters. Our automatic approach selects a parameter

setting among those with best performance.

In addition we compute the area under the curve

(AUC) for Recall and Precision as in (Min et al.,

2004). The AUC is a comparison measure that is in-

dependent of a particular overlap threshold. A perfect

system would have AUC of 1.0.

4.2 Results and Interpretation

The experiments are evaluated on public benchmark

data (Fisher, 2004). These are 27 video sequences

with hand labelled ground truth data. We divided

this set in 13 sequences for training (18411 bounding

boxes) and 14 sequences for testing (21217 boxes).

Quality of the automatic parameter regulation

This experiment gives insight in the quality of the au-

tomatically selected parameter setting using our ap-

proach. For given sequence, our approach selects

a parameter setting using a search strategy by a ge-

netic algorithm. For parameter selection we use

the generated scene reference model with the metric

avg

(y(t)). This parameter setting is then used to

generate an output in XML format. The statistics tool

provides values for recall and precision for an over-

lap threshold of T =0.5. This recall/precision pair

is represented as a point in the precision recall plot of

Figure 3.

To demonstrate the quality of this result, we com-

pare it to the precision recall values of a large number

of different parameter settings (all parameter settings

that are selected by the genetic algorithm, 3 genera-

tions of 20 individuals). Figure 3 shows clearly, that

the automatically selected parameter setting is among

the settings with the best performance. For other se-

quences we observe a similar behaviour.

Performance increase by auto-regulation

We measure the performance of the tracking system

VISAPP 2006 - IMAGE UNDERSTANDING

Table 1: Performance (on 14 sequences) of self-adaptive

tracking and benchmarks.

Method AUC

recall

AUC

precision

Manual tuning 0.429 0.532

(on 6 sequences)

Auto-regulation 0.437 0.629

using KM-F (batch)

Auto-regulation 0.431 0.648

using Histo-EM-F

No regulation 0.414 0.391

(low thresholds)

No regulation 0.243 0.426

(high thresholds)

by evaluating the output of the system that uses the

parameter setting chosen by the auto-regulation algo-

rithm. Table 1 shows the results of the different meth-

ods on the 14 test sequences. The auto-regulation

methods are compared to three benchmarks. The up-

per benchmark is our tracking system with manually

tuned parameters. We only have the results for 6 test

sequences due to the tedious manual task. The lower

benchmarks are provided by using static parameter

settings for all test sequences. Low thresholds give

good recall but bad precision, high thresholds give

bad recall and bad precision.

We compare the auto-regulation method using a

batch model KM-F and the best performing incremen-

tal method Histo-EM-F. The KM-F model is obtained

using a KMeans clustering with 1000 initial clusters

that are subsequently merged. All clustering solutions

from 100 to 5 clusters are transformed to a GMM.

The model obtained by the method Histo-EM-F uses

a ﬁne grain histogram for initialisation of EM. The

Gaussians computed by EM are subsequently merged

to provide a fusion tree. In both methods, we choose

the least complex model with a P

err

value below the

acceptance threshold of 5% (see paragraph below).

The incremental model and the batch model both

match the performance of a manually tuned system

(although the manually tuned system is evaluated only

on 6 sequences, the AUC values give an idea of the

performance range). The increase in performance us-

ing automatic parameter regulation with respect to us-

ing no regulation is demonstrated clearly by the com-

parison to the lower benchmark performance. An-

other important result is that no signiﬁcant differ-

ence can be noted between the incremental and non-

incremental model. This motivates the use of incre-

mental models in the future due to their ability of fur-

ther reﬁnement with additional data.

Link between P

err

and tracking performance

In section 3.2 we proposed a measure for model se-

lection based on the probability of classiﬁcation er-

ror. This is a convenient and fast measure for model

selection. The deﬁnition and validation of such a fast

quality measure makes possible the automatic gener-

ation and selection of a scene reference model in a

non-supervised commercial tracking application. For

validation of this measure, we need to show the rela-

tion between P

err

and the tracking performance.

We perform following experiment. 4 sequences are

selected among the test sequences with different de-

gree of difﬁculty (1 easy, 2 intermediate and 1 hard).

We compute the P

err

scores for a family of mod-

els extracted from different levels of the fusion tree

produced by the GNGN-EM-F approach (a growing

neural gas network (Fritzke, 1995) is used for initial-

isation of EM with fusion). P

err

requires the selec-

tion of a representative set of positive and negative

examples. The positive examples are composed of 7

trajectories of the most common paths of the scene

(2420 measurements). The negative examples are ob-

tained by monitoring the tracking output and collect-

ing tracking errors (1997 measurements). Each model

is then used for auto-regulation of parameters. The

XML output of the tracking system using the selected

parameter setting is evaluated and the AUC values of

precision and recall are computed.

Table 2 shows these results. The results conﬁrm the

relation between the P

err

value and the global per-

formance of the system using auto-regulation. Small

err

values (good model quality) yield good track-

ing performance (high precision and high recall). The

results show also that no further improvement is ob-

served for P

err

values below 5%. We observed a sim-

ilar behaviour for model families obtained by other

learning approaches. The performance stabilises for

err

values below 5%. As a consequence, in this par-

ticular system setup, a model with a P

err

value below

5% is appropriate for parameter regulation.

5 CONCLUSION AND OUTLOOK

We described an architecture for a self-adaptive

tracking system that uses a control component to

implement the abilities of auto-criticism and auto-

regulation. Both modules require a metric with re-

spect to a reference model. Our approach allows to

automatically generate and select such a reference

model with good quality without human supervision.

The quality of the model is determined by a fast eval-

uation measure based on the classiﬁcation error with

respect to a validation data set.

The experiments show that a tracking system with

auto-regulation of parameters has the same or better

performance than a tracking system with manually

tuned parameters. We also demonstrate that our au-

tomatic parameter selection scheme selects parame-

ter settings with very high performance. The exper-

AN AUTOMATIC APPROACH FOR PARAMETER SELECTION IN SELF-ADAPTIVE TRACKING

Table 2: Performance (evaluated on 4 sequences) of self-adaptive system using incremental models with increasing P

err

Performance degrades as P

err

increases.

err

(in %) 3.33 4.04 4.26 5.34 7.04 9.16 11.7 14.9

AUC

recall

0.47 0.47 0.47 0.46 0.45 0.40 0.44 0.38

AUC

precision

0.52 0.56 0.55 0.58 0.58 0.48 0.29 0.39

iments validate the fast evaluation measure P

err

example sequences.

With respect to a future integration in a commer-

cial tracking system, the result that incremental model

generation produces equal results than batch model

generation is important. Incremental learning tech-

niques have the advantage that they allow subsequent

reﬁnement of the reference model without the need

of storing all training data. This is a great advan-

tage with respect to direct learning methods such as

k-means clustering or EM whose computational com-

plexity depends on the total number training data. The

data required for a model with very good quality may

reach quickly the upper limit of memory space and

available computation time.

The next step of our work consists in developing

a prototype that can be used for an automatic in-

stallation of a self-adaptive tracking system in a new

site. The here proposed technique is fully automatic

once the ground truth data is acquired. Hand-labelling

ground truth may be replaced by a robust and reli-

able tracker using colour information. An initialisa-

tion phase would require the cooperation of several

actors wearing coloured suits. 10 minutes of acting

provides 18000 frames of example data which are suf-

ﬁcient for a scene reference model of a lobby with

several entries and unconstrained walking paths. Fol-

lowing these steps, a scene reference model for a new

video surveillance site can be generated automatically

using the here proposed approach.

REFERENCES

Caporossi, A., Hall, D., Reignier, P., and Crowley, J.

(2004). Robust visual tracking from dynamic control

of processing. In International Workshop on Perfor-

mance Evaluation of Tracking anS

urveillance, pages

23–31.

Fisher, R. (2004). The PETS04 surveillance ground-truth

data sets. In International Workshop on Performance

Evaluation of Tracking anS

urveillance.

Fritzke, B. (1995). A growing neural gas network learns

topologies. In Tesauro, G., Touretzky, D., and Leen,

T., editors, Advances in Neural Information Process-

ing Systems, volume 7, pages 625–632. MIT Press,

Cambridge, MA, USA.

Hall, D. (2005). Automatic parameter regulation for a track-

ing system with an auto-critical function. In Inter-

national Workshop on Computer Architecture for Ma-

chine Perception, pages 39–45.

Makris, D. and Ellis, T. (2003). Automatic learning of an

activity-based semantic scene model. In International

Conference on Advanced Video and Signal Based Sur-

veillance, pages 183–188.

Min, J., Powell, M., and Bowyer, K. (2004). Automated

performance evaluation of range image segmentation

algorithms. IEEE Transactions on Systems Man and

Cybernetics, 34(1):263–271.

Murino, V., Foresti, G., and Regazzoni, C. (1996). A dis-

tributed probabilistic system for adaptive regulation of

image processing parameters. IEEE Transactions on

systems, man, and cybernetics, 26(1):1–20.

Robertson, P. and Brady, J. (1999). Adaptive image analy-

sis for aerial surveillance. IEEE Intelligent Systems,

14(3):30–36.

Scotti, G., Marcenaro, L., and Regazzoni, C. (2003). A

s.o.m. based algorithm for video surveillance system

parameter optimal selection. In IEEE Conference on

Advanced Video and Signal Based Surveillance.

VISAPP 2006 - IMAGE UNDERSTANDING