See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228873288

Performance evaluation of object tracking algorithms

Article · January 2007

CITATIONS READS

123 12,928

3 authors, including:

Dimitrios Makris Sergio Velastin

Kingston University University Carlos III de Madrid
119 PUBLICATIONS 3,237 CITATIONS 293 PUBLICATIONS 8,243 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Dimitrios Makris on 23 May 2014.

The user has requested enhancement of the downloaded file.

 Performance Evaluation of Object Tracking Algorithms

Fei Yin, Dimitrios Makris, Sergio Velastin

Digital Imaging Research Centre, Kingston University London, UK
{fei.yin, d.makris, sergio.velastin}@kingston.ac.uk

Abstract detection, based on ROC-like curves and the F-

This paper deals with the non-trivial problem of measure. The latter allows straight-forward comparison
performance evaluation of motion tracking. We using a single value that takes into account the
propose a rich set of metrics to assess different aspects application domain.
of performance of motion tracking. We use six different Needham and Boyle [9] proposed a set of metrics
video sequences that represent a variety of challenges and statistics for comparing trajectories and evaluating
to illustrate the practical value of the proposed metrics tracking motion systems.
by evaluating and comparing two motion tracking Brown et al [10] suggest a motion tracking
algorithms. The contribution of our framework is that evaluation framework that estimates the number of
allows the identification of specific weaknesses of True Positive (TP), False Positive (FP) and False
motion trackers, such as the performance of specific Negative (FN), Merged and Split trajectories. However
modules or failures under specific conditions. their definition, based on the comparison of the system
track centroid and an enlarged ground truth bounding
1. Introduction box) favours tracks of large objects.
Bashir and Porikli [11] gave definitions of the above
Significant research effort has focused on video- metrics based on the spatial overlap of ground truth and
based motion tracking [1] [2] [3] [4] and attract the system bounding boxes that are not biased towards
interest of industry. Performance evaluation of motion large objects. However they are counted in terms of
tracking is important not only for the comparison and frame samples. Such an approach is justified when the
further development of algorithms from researchers, objective of performance evaluation is object detection
but also for the commercialisation and standardisation [7] [8]. In object tracking, measuring TP, FP and FN in
of the technology as typified by the i-LIDS Programme terms of tracks rather than frames is a natural choice
in the UK [5]. that is consistent to the expectations of the end-users.
In this paper, we have selected a set of motion This paper is organized as follows: Section 2
tracking metrics that are used to highlight different defines provides the definitions for motion tracking and
aspects of motion tracking performance. We illustrate track. Section 3 describes the performance evaluation
the purpose of the proposed metrics in the evaluation of methodology and gives definition of different metrics.
two motion tracking algorithms, using a variety of Results are presented and discussed in section 4.
datasets. Section 5 concludes the paper.
Ellis [6] investigated the main requirements for
effective performance analysis for surveillance systems 2. Motion Tracking
and proposed some methods for characterising video
datasets. We define motion tracking as the problem of
Nascimento and Marques [7] proposed a framework estimating the position and the spatial extent of the
which compares the output of different motion non-background objects for each frame of a video
detection algorithms against given ground truth and sequence. The result of motion tracking is a set of
estimates objective metrics such as Correct Detections, tracks Tj, j=1..M, for all M moving objects of the
False alarms, Detection failures, Merges and Splits. scene. A track Tj is defined as: Tj={xij, Bij}, i =1..N,
They also proposed ROC-like curves that can where xij and Bij are the centre and the spatial extent
characterize algorithms over a range of parameters. (usually represented by a bounding box) respectively of
Lazarevic-McManus et al [8] developed an object- the object j for the frame i and N is the number of
based approach to enable evaluation of motion frames.
 where Length is the number of frames and TRov an
3 Performance Evaluation arbitrary threshold. If Eq.4 is true, then, we associate
the system track with the GT track and start evaluating
3.1 Preparation the performance of the system track.

We propose a set of metrics that compare the output

of motion tracking systems to a Ground Truth in order
to evaluate the performance of the systems.
Before the evaluation metrics are introduced, it is
important to define the concepts of spatial and
temporal overlap between tracks, which are required
to quantify the level of matching between Ground
Truth (GT) tracks and System (ST) tracks, both in
Figure 2 Example of track overlapping
space and time.
The spatial overlap is defined as the overlapping
level A(GTi, STj) between GTi and STj tracks in a
specific frame k (Fig. 1). 3.2 Metrics
Area (GTik I ST jk )
A(GTik , ST jk ) = (1) In this section, we give definitions of high level
Area (GTik U ST jk ) metrics such as True Positive (TP), False Positive (FP)
and False Negative (FN) tracks. Such metrics are useful
because they are the base for estimating metrics such as
GTik GTik Specificity and Accuracy [11] and allow the
construction of ROC-like curves [8]. Metrics such as
STjk STjk
Track Fragmentation and ID Change assess the
integrity of tracks. Finally, we define metrics that
measure the accuracy of motion tracking in estimating
Figure 1: Area (GTik I ST jk ) and Area (GTik U ST jk ) the position (Track Matching Error), the spatial extent
(Closeness), the completeness and the temporal
We also define the binary variable O(GTi, STj), latency.
based on a threshold Tov which in our examples is
arbitrarily set to 20%. Correct detected track (CDT) or True Positive
(TP):
1 if A(GTik , ST jk ) > Tov A GT track is considered to have been detected
O(GTik , ST jk ) =  (2) correctly if it satisfies both of the following conditions:
0 if A(GTik , ST jk ) ≤ Tov
Condition 1: The temporal overlap between GT
Track i and system track j is larger than a predefined
Temporal overlap TO(GTi, STj) is a number that track overlap threshold TROV which in our examples is
indicates overlap of frame span between system track j arbitrarily set to 15%.
and GT track i: Length (GTi I ST j )
>= TRov (5)
TO − TO S , TO E > TO S Length(GTi )
TO (GTi , ST j ) =  E (3)
 0, TO E ≤ TO S Condition 2: The system track j has sufficient
where TOS is the maximum of the first frame indexes spatial overlap with GT track i.
of TOE is the minimum of the last frame indexes of the N

two tracks. The temporal and spatial overlap between ∑ A(GT

k =1
ik , ST jk )
tracks is illustrated graphically in Fig. 2: >= Tov (6)
We use a temporal-overlap criterion to associate N
systems tracks to GT tracks according to the following Each GT track is compared to all system tracks
condition: according to the conditions above. Even if there is
Length (GTi I ST j ) more than one system track meets the conditions for
> TRov (4) one GT track (which is probably due to
Length (GTi U ST j ) fragmentation), we still consider the GT track to have
been correctly detected. Fragmentation errors are
counted separately (see below). So, if each of the GT Condition 2: A GT track i does not have any sufficient
tracks is detected correctly (by one or more system spatial overlap with any system track, although it has
tracks), the number of CDT equals the number of GT enough temporal overlap with system track j.
tracks. N

∑ A(GT ik , ST jk )
False alarm track (FAT) of False Positive (FP):
k =1
< Tov (10)
N
Although it is easy for human operators to realise
what is a false alarm track (event) even in complex
Similar definitions of the above metrics have been
situation, it is hard for an automated system to do so.
given in [10] and [11]. In [10], spatial overlap is
Here, we give a practical definition of false alarm track
defined by checking whether the centroid of the system
(Fig. 3). We consider a system track as false alarm if
track is within the area of the GT track, enlarged by
the system track meets any of the following conditions:
20%. However, such a definition favours tracks of
Condition 1: A system track j does not have
large objects. In [11], spatial overlap is defined in three
temporal overlap larger than TROV with any GT track i.
different ways: a) Euclidean distance between
Length(GTi I ST j )
< TRov (7) centroids, b) system track centroid within GT area and
Length( ST j ) c) area ratio overlap. Our approach is similar to the
latter definition. However, [11] estimates TP, FP, and
Condition 2: A system track j does not have FN in terms of frame samples, while we think that they
sufficient spatial overlap with any GT track although it must be measured in terms of tracks (using the concept
has enough temporal overlap with GT track i. of temporal overlap) to be more informative to the end-
N user.
∑ A(GT ik , ST jk )
k =1
< Tov (8) Track Fragmentation (TF):
N Fragmentation indicates the lack of continuity of
system track for a single GT track. Fig. 4 shows an
example of track fragmentation error:

Figure 4: The number of track fragmentations is

TF=2 (the system track fragmented two times).
Figure 3 Example of two false alarm tracks. The
In an optimal condition, track fragmentation error
left ST fails condition (Eq.7), while the right ST
should be zero which means the tracking system is able
fails condition (Eq.8)
to produce continuous and stable tracking for the
ground truth object.
FAT is an important metric because it is
As mentioned before, we allow multiple
consistently indicated by operators that a system which
associations between GT track and system track
does not have a false alarm rate close to zero is likely
therefore fragmentation is measured from the track
to be switched off, not matter its TP performance.
correspondence results.
Track detection failure (TDF):
A GT track is considered to have not been detected ID Change (IDC):
(i.e. it is classified as a track detection failure), if it We introduce the metric IDCj to count the number
satisfies any of the following conditions. of ID changes for each STj track. Note that such a
Condition 1: A GT track i does not have temporal metric provides more elementary information than an
overlap larger than TROV with any system track j. ID swap metric.
For each frame k, the bounding box Dj,k of the
Length (GTi I ST j )
< TRov (9) system track STj may be overlapped with N Dj ,k GT
Length(GTi )
areas, where N Dj ,k is given by:
N Dj ,k = ∑ O(Gik , D jk ) (11)
i
 We take into account only the frames for which enough to trigger the tracking in time or indicates that
N Dj ,k =1 (which means that the track STj is associated the detection is not good enough to trigger the tracking.
It is estimated by the difference in frames between
(spatially overlapped) with only one GT Track for each the first frame of system track and the first frame of GT
of these frames). We use these frames to estimate the track.
ID changes of STj as the number of changes of (13)
LT = start frame of ST j − start frame of GTi
associated GT tracks.
We can estimate the total number of IDC changes
in a video sequences as:
IDC = ∑ IDC j (12)
j
The procedure for counting ID change is shown in Fig.
5. Some examples of estimating ID changes are shown
in Fig. 6:
Figure 7: Example of Latency of system track
For each System track j;
IDCj=-1 Closeness of Track (CT):
For every frame k; For a pair of associated GT track and system track,
If there is no occlusion, and N Dj ,k =1; a sequence of spatial overlaps (Fig. 2) is estimated by
If IDCj=-1 or M(IDCj) ≠i Eq.2 for the period of temporal overlap:
( i is the ID of GT track whose area is CT (GTi , ST j ) = { A(GTi1 , ST j1 ),... A(GTiN ES , ST jN ES )} (14)
overlapped with the GT track in this frame.) From Eq.14, we can estimate the average closeness
Then,
IDCj=IDCj+1; M(IDCj) = i;
for the specific pair of GT and system tracks. To
End compare all M pairs in one video sequence, we define
IDC=sum(IDCj) the closeness of this video as the weighted average of
End track closeness of all M pairs:
M
End
Figure 5: Pseudo-code for estimated ID Changes
∑ CT
t =1
t
(15)
CTM =
(IDC) M

∑ Length (CT )
t =1
t

and the weighted standard deviation of track closeness

for the whole sequence:
M

∑ Length(CT ) × std (CT )

t =1
t t
(16)
CTD = M
,
ID swap = 2 ID changes No ID change
∑ Length(CT )
t =1
t

where std(CTt) is the standard deviation of CTt

Track Matching Error (TME):

ID change No ID change This metric measures the positional error of system
tracks. Fig. 8 shows positions of a pair of tracks.
Figure 6: Examples of ID swap and ID change

Latency of the system track (LT):

Latency (time delay) of the system track is the time
gap between the time that an object starts to be tracked Figure 8 Example of a pair of trajectories
by the system and the first appearance of the object
(Fig.7). The optimal latency should be zero. A very TME is the average distance error between a system
large latency means the system may not be sensitive track and the GT track. The smaller the TME, the
better the accuracy of the system track will be.
 N I

∑ Dist (GTC , STC ) ik jk ∑ {max(TC

i =1
GTi ) − TCT }
(22)
TME = k =1 (17) TCD =
Length(GT I ST ) i j
N −1

where Dist() is the Euclidean distance between the 4. Results

centroids of GT and the system track:
We demonstrate the practical value of the proposed
TMED is the standard deviation of distance errors, metrics by evaluating two motion tracking systems (an
which is defined as: experimental industrial tracker from BARCO and the
2
OpenCV1.0 blobtracker [12]). We run the trackers on
N

∑ ( Dist (GTCik , STC jk ) − TMEM ) six video sequences (shown in Fig.9-Fig.14) that
represent a variety of challenges, such as illumination
TMED = k =1 (18)
Length(GTi I ST j ) − 1 changes, shadows, snow storm, quick moving objects,
blurring of FOV, slow moving objects, mirror image of
Similarly, track matching error (TMEMT) for the objects and multiple object intersections. The ground
whole video sequence is defined as the weighted truth for all videos was manually generated using Viper
average over the duration of overlapping of each pair GT [13].
of tracks as the weight coefficient.
M

∑ Length(GT
t =1
i I ST j )t × TMEt
(19)
TMEMT = M

∑ Length(GT
t =1
i I ST j )t

and the standard deviation of track matching errors for

the whole sequence: Figure 9: PETS2001 PetsD1TeC1.avi sequence is
M 2686 frames (00:01:29) long and depicts 4
∑ Length(GT I ST ) × TMED
t =1
i j t t
(20)
persons, 2 groups of persons and 3 vehicles. Its
TMEMTD = M
main challenge is the multiple object intersections.
∑ Length(GT I ST )
t =1
i j t

Track Completeness (TC):

This is defined as the time span that the system
track overlapped with GT track divided by the total
time span of GT track. A fully complete track is where
this value is 100%. Figure 10: i-LIDS SZTRA103b15.mov sequence is
N ES
5821 frames (00:03:52) long and depicts 1
∑ O(GT ik , ST jk )
person. Its main challenges are the illuminations
TC = k =1 (21)
number of GTi changes and a quick moving object.

If there is more than one system track associated

with the GT track, then we choose the maximum
completeness for each GT track
Also, we define the average track completeness of a
video sequence as:
I

∑ max(TC GTi ) Figure 11: i-LIDS SZTRA104a02.mov sequence is

TCM = i =1 (22) 4299 frames (00:02:52) long and depicts one
N person.
where N is the number of GT tracks and the standard
deviation of track completeness for the whole sequence
is:
 Table 1 Evaluation results for PETS2001 Seq.
PetsD1TeC1.avi BARCO OpenCV
tracker tracker
GT Tracks 9 9
System Tracks 12 17
CDT 9 9
Figure 12: i-LIDS PVTRA301b04.mov sequence is FAT 3 6
7309 frames (00:04:52) long and depicts 12 TDF 0 0
persons and 90 vehicles. Its main challenges are TF 3 3
shadows, moving object in the beginning of IDC 5 7
sequence and multiple object intersections. LT 46 66
CTM 0.47 0.44
CTD 0.24 0.14
TMEMT 15.75 5.79
TMEMTD 23.64 5.27
TCM 0.67 0.58
TCD 0.24 0.89

Figure 13: BARCO 060306_04_Parkingstab.avi is Table 2 BARCO track association results for
7001 frames long and depicts 3 pedestrians and 1 Pets2001
vehicle. Its main challenge is the quick illumination
PetsD1TeC1.avi: BARCO tracker
changes.
GT-ID ST-ID LT CTM CTD TC
0 1 198 0.571 3.19 0.61
1 2 27 0.677 8.34 0.91

2 3 49 0.512 11.64 0.29

2 5 341 0.410 28.90 0.55
3 1 26 0.529 14.64 0.15
3 5 252 0.256 54.46 0.08
Figure 14: BARCO 060306_02_Snowdivx.avi is 4 4 35 0.639 5.91 0.84
8001 frames long and depicts 3 pedestrians. Its 5 5 0 0.391 10.80 0.62
main challenges are snow storm, blurring of FOV, 6 5 0 0.049 184.60 0.14
slow moving objects and mirror image of objects. 6 4 95 0.528 7.21 0.77
7 1 32 0.460 5.83 0.88
The results of performance evaluation are presented 8 5 0 0.420 18.80 0.87
in Tables 1-8. Generally, high level metrics such CDT,
FAT, TDF show that the BARCO tracker outperforms Table 3 OpenCV track association results for
the OpenCV tracker in almost all cases. The only PETS2001
exception is the i-Lids sequence PVTRA301b04.mov PetsD1TeC1.avi: OpenCV tracker
(Fig. 12, Table 6). On the other hand, the OpenCV GTID cvSTID LT CTM CTD TC
tracker is generally more accurate in estimating the 0 1 16 0.54 2.13 0.97
position of the objects (lower TMEM). 1 2 56 0.35 11.83 0.82
Also, the OpenCV tracker dealt better with the snow
2 7 24 0.46 8.95 0.27
scene (Fig.14, Table8) in estimating the position and
the spatial and temporal extent of the objects (lower 2 12 0 0.49 5.01 0.13
TMEM, higher CTM and TCM), which implies a better 3 3 226 0.36 9.89 0.09
object segmentation module for this scene. However, 3 5 346 0.56 4.72 0.12
the BARCO tracker has better high level metrics (lower 4 4 120 0.52 6.24 0.56
FAT, lower TF), which implies a better tracking policy. 5 6 166 0.50 7.37 0.33
Note that without the rich set of metrics as used here 6 11 12 0.40 3.60 0.98
it is very difficult to identify possible causes of 7 14 26 0.37 2.87 0.81
poor/good performance in different trackers. 8 13 14 0.45 10.16 0.40
8 15 300 0.34 15.59 0.28
 LT 57 78
Table 4 Tracking PE results for i-LIDS CTM 0.30 0.12
SZTRA103b15 CTD 0.21 0.16
SZTRA103b15 BARCO OpenCV TMEMT 49.70 24.65
.mov tracker tracker TMEMTD 60.31 22.85
GT Tracks 1 1 TCM 0.34 0.26
8 TCD 0.57 0.65
System Tracks 15
CDT 1 1
Table 7 Tracking PE results for BARCO
FAT 3 12 Parkingstab
TDF 0 1 Parkingstab.avi BARCO OpenCV
TF 0 0 tracker tracker
IDC 0 0 GT Tracks 4 4
LT 50 9 System Tracks 9 17
CTM 0.65 0.23 CDT 4 4
CTD 0.21 0.10 FAT 1 11
TMEMT 9.10 15.05 TDF 0 0
TMEMTD 12.48 3.04 TF 0 0
TCM 0.68 0.42 IDC 0 0
TCD 0.00 0.00 LT 72 35
CTM 0.50 0.39
Table 5 Tracking PE results for i-LIDS CTD 0.20 0.14
SZTRA104a02 TMEMT 13.32 11.82
SZTRA104a02 BARCO OpenCV TMEMTD 11.55 8.16
.mov tracker tracker TCM 0.82 0.77
TCD 0.11 0.96
GT Tracks 1 1
System Tracks 4 9
CDT 1 1 Table 8 Tracking PE results for BARCO Snowdivx
FAT 0 5 Snowdivx.avi BARCO OpenCV
TDF 0 0 tracker tracker
TF 0 2 GT Tracks 3 3
IDC 0 0 System Tracks 28 29
LT 74 32 CDT 3 3
CTM 0.79 0.34 FAT 19 20
CTD 0.21 0.17 TDF 0 0
TMEMT 7.02 16.69 TF 2 5
TMEMTD 15.67 7.55 IDC 0 0
TCM 0.73 0.44 LT 590 222
TCD 0.00 0.00 CTM 0.14 0.42
CTD 0.23 0.12
Table 6 Tracking PE results for i-LIDS TMEMT 28.50 16.69
PVTRA301b04 TMEMTD 35.44 11.62
PVTRA301b04 BARCO OpenCV TCM 0.33 0.35
.mov tracker tracker TCD 0.47 0.71
GT Tracks 102 102
System Tracks 225 362 5. Conclusions
CDT 90 95
FAT 67 112 We presented a new set of metrics to assess
TDF 12 7 different aspects of performance of motion tracking.
TF 62 98 We proposed statistical metrics, such as Track
IDC 95 102 matching Error (TME), Closeness of Tracks (CT) and
 Track Completeness (TC) that indicate the accuracy of Imagery library for intelligent detection systems (i-LIDS),
estimating the position, the spatial and temporal extent http://scienceandresearch.homeoffice.gov.uk/hosdb/cctv-
of the objects respectively and they are closely related imaging-technology/video-based-detection-systems/i-lids/,
[Last accessed: August 2007]
to the motion segmentation module of the tracker.
Metrics, such as Correct Detection Track (CDT), [6] T. Ellis, “Performance Metrics and Methods for
False Alarm Track (FAT) and Track Detection Failure Tracking in Surveillance”, Third IEEE International
(TDF) provide a general overview of the algorithm Workshop on Performance Evaluation of Tracking and
performance. Track Fragmentation (TF) shows the Surveillance, June, Copenhagen, Denmark, 2002, pp26-31.
temporal coherence of tracks. ID Change (IDC) is [7] J. Nascimento, J. Marques, “Performance evaluation of
useful to test the data association module of multi- object detection algorithms for video surveillance”, IEEE
target trackers. Transactions on Multimedia, 2005, pp761-774.
We tested two trackers using six video sequences [8] N.Lazarevic-McManus, J.R.Renno, D. Makris,
that provide a variety of challenges, such as G.A.Jones, “An Object-based Comparative Methodology for
illumination changes, shadows, snow storm, quick Motion Detection based on the F-Measure”, in 'Computer
moving objects, blurring of FOV, slow moving objects, Vision and Image Understanding', Special Issue on
mirror image of objects and multiple object Intelligent Visual Surveillance TO APPEAR, 2007.
intersections. [9] C.J. Needham, R.D. Boyle. “Performance Evaluation
The variety of metrics and datasets allows us to Metrics and Statistics for Positional Tracker Evaluation”
reason about the weaknesses of particular modules of International Conference on Computer Vision Systems
the trackers against specific challenges, assuming (ICVS'03), Graz, Austria, April 2003, pp. 278 - 289.
orthogonality of modules and challenges. This
[10] L. M. Brown, A. W. Senior, Ying-li Tian, Jonathan
approach is a realistic way to understand the drawbacks Connell, Arun Hampapur, Chiao-Fe Shu, Hans Merkl, Max
of motion trackers, which is important for improving Lu, “Performance Evaluation of Surveillance Systems Under
them. Varying Conditions”, IEEE Int'l Workshop on Performance
In future work, we will use this framework for Evaluation of Tracking and Surveillance, Colorado, Jan
evaluating more trackers. We will also extend the 2005,.
framework to allow evaluation of high level tasks such [11] F. Bashir, F. Porikli. “Performance evaluation of object
as event detection and action recognition. detection and tracking systems”, IEEE International
Workshop on Performance Evaluation of Tracking and
6. Acknowledgments Surveillance (PETS), June 2006
[12] OpenCV Computer Vision Library
The authors would like to acknowledge financial http://www.intel.com/technology/computing/opencv/index.ht
support from BARCO View, Belgium and the m, [Last accessed: August 2007]
Engineering and Physical Sciences Research Council
[13] Guide to Authoring Media Ground Truth with ViPER-
(EPSRC) REASON project under grant number GT, http://viper-toolkit.sourceforge.net/docs/gt/, [Last
EP/C533410. accessed: July 2007]
[14] Pets Metrics, http://www.petsmetrics.net/ [Last
7. References accessed: August 2007]
[1] I. Haritaoglu, D. Harwood, and L. S. Davis, “W : Real-
time surveillance of people and their activities,” IEEE Trans.
Pattern Anal. Machine Intell., Aug. 2000, pp. 809 - 830
[2] M. Isard and A. Blake, “Contour tracking by stochastic
propagation of conditional density,” in Proc. European Conf.
Computer Vision, 1996, pp. 343 - 356
[3] M. Xu, T.J. Ellis, “Partial observation vs. blind
tracking through occlusion”, British Machine Vision
Conference, BMVA, September, Cardiff, 2002, pp. 777 -786
[4] Comaniciu D., Ramesh V. and Meer P., \Kernel-Based
Object Tracking", IEEE Trans. Pattern Anal. Machine Intell.,
2003. pp. 564 - 575
[5] Home Office Scientific Development Branch

View publication stats