A Method for Detecting Long Term Left Baggage based on Heat Map

Pasquale Foggia, Antonio Greco, Alessia Saggese and Mario Vento

∗

Dept. of Computer Eng. and Electrical Eng. and Applied Mathematics, University of Salerno,

Via Giovanni Paolo II, 132, Fisciano (SA), Italy

Keywords:

Left Bag, Abandoned Luggage, Stopped Object Detection.

Abstract:

In this paper we propose a method able to identify the presence of objects remaining motionless in the scene

for a long time by analyzing the videos acquired by surveillance cameras. Our approach combines a back-

ground subtraction strategy with an enhanced tracking algorithm. The main contributions of this paper are

the following: ﬁrst, spatio-temporal information is implicitly encoded into a heat map; furthermore, differ-

ently from state of the art methodologies, the background is not updated by only evaluating the instantaneous

movement of the objects, but instead by taking into account their whole history encoded in the heat map. The

experimentation has been carried out over two standard datasets and the obtained results have been compared

with state of the art approaches, conﬁrming the effectiveness and the robustness of our system.

1 INTRODUCTION

In the last years the research community has shown a

great interest toward the problem of detecting stopped

objects. This is mainly due to the wide range of ap-

plicative ﬁelds where such technology may be prof-

itably used, ranging from left baggages detection in

metro stations and airports to the detection of garbage

illegally dumped along the streets.

Although it is not possible to partition the ex-

isting methods into clear-cut, non-overlapping cat-

egories, two different typologies of approaches can

be identiﬁed, namely tracking-based and background-

subtraction based. The methods belonging to the for-

mer category ﬁrst locate a foreground object when

it is moving and then analyze its trajectory so as

to check if it becomes a stationary object. On the

other hand, the approaches based on the second cate-

gory take advantage of a properly deﬁned background

model and on foreground extraction techniques to de-

tect stopped objects.

As for the tracking-based approaches, in (Guler

et al., 2007) a 4-level tracking method inspired by the

human visual attention model is proposed: the con-

sidered levels are peripheral tracker, vision tunnels,

scene description layer and stationary object layer.

An object is considered stationary if its dwell time

within the same region exceeds a given threshold,

chosen by the human operator during the conﬁgu-

∗

IAPR Fellow

ration step. In (Bhargava et al., 2007) the authors

propose a backtracking approach to detect abandoned

luggage in crowded scenes. The attention is not only

focused on baggage, but also on the owner: in fact,

the system detects a stopped bag and performs back-

tracking to keep track of the owner, generating an

alarm only if he does not retrieve the luggage within

60 seconds. The joint analysis of the bag and its

owner is also exploited in (Acampora et al., 2012),

where moving objects trajectories are analyzed by a

Time Delay Neural Network and a decision about the

event of interest is taken by exploiting a set of man-

ually deﬁned fuzzy rules. In (Bevilacqua and Vac-

cari, 2007) the authors focus on the problem of sta-

tionary vehicles detection, by proposing a tracking al-

gorithm based on corner point detection. Occlusions

are solved with a SOM neural network and the tra-

jectories are smoothed with a moving average, so as

to partially eliminate noise caused by tracking algo-

rithms and to facilitate stopped delay measurement:

for each vehicle, its trajectory is analyzed; if it is sta-

ble in a certain region and for a long time interval,

then it is considered stopped.

The above mentioned algorithms are very intu-

itive and achieve a very high performance in ster-

ile environments, where only a few people populate

the scene. However, their main limitation lies in the

fact that they cannot be effectively used in crowded

scenes: in fact, they are very sensitive to occlusions,

which usually prevent the system from tracking ob-

385

Foggia P., Greco A., Saggese A. and Vento M..

A Method for Detecting Long Term Left Baggage based on Heat Map.

DOI: 10.5220/0005306803850391

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 385-391

ISBN: 978-989-758-090-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

jects and then from extracting trajectories in a reli-

able way. On the other hand, methods based on back-

ground subtraction are able to overcome such limi-

tation and in general to achieve better performance.

This is due to the fact that these methods perform a

temporal analysis of foreground objects by sophisti-

cated background modeling strategies, and thus com-

mon problems such as occlusions cannot inﬂuence the

performance of this kind of systems.

For instance, the authors in (Maddalena and Pet-

rosino, 2013) propose a general framework, called

stopped foreground subtraction (SFS), that is inde-

pendent of the speciﬁc background modeling and

foreground extraction methods: in fact, they created

a model of the stopped objects and use it to clas-

sify a new stationary one. Furthermore, they propose

a background updating method based on neural net-

works, called 3DSOBS, which proved its effective-

ness if compared with the traditional MOG (Stauf-

fer and Grimson, 1999). In (Porikli et al., 2008) the

algorithm maintains two different backgrounds, up-

dated with different speeds by using the traditional

MOG: a short term and a long term one. Thus, an ev-

idence image is computed by considering the pixels

whose change rate ranges between the short term and

the long term background updating rate; this image

is ﬁnally used to detect stopped objects. The method

proposed in (Boragno et al., 2007) is developed on

the Ipsotek VI platform and is able to detect vehicles

parked in prohibited areas. The algorithm is based

on 3 steps: motion detection through block matching,

stopped object detection using MOG and object clas-

siﬁcation into trucks, cars, pedestrians and packages.

Starting from the spatio-temporal information ob-

tained from the background analysis, several methods

also include a tracking phase in order to extract more

useful information about the objects in the scene, so

as to increase the detection capabilities of the system

as well as to decrease their false alarm rate. Such al-

gorithms are usually referred to as hybrid approaches

For instance in (Smitha and Palanisamy, 2012) a

parked vehicles detection algorithm is described; it

uses a simple background subtraction technique and

a region-based tracking to identify the stop of a ve-

hicle. The method proposed in (Singh et al., 2009)

is based on a dual background subtraction technique

to detect stationary objects and on a tracking algo-

rithm, optimized to solve occlusions, able to reduce

the number of false positives introduced by the pure

background subtraction approach. In fact, the sys-

tem raises an alarm only if the object hit count ex-

ceeds a given threshold, i.e. if it remains stationary

for a certain time interval. In (Venetianer et al., 2007)

the authors propose a method for abandoned objects

or stopped vehicles detection developed on the Ob-

jectVideo commercial platform. The algorithm con-

sists of 4 steps: background subtraction, blob detec-

tion, tracking and stopped object detection. The last

step is performed bringing almost immediately the

stationary object in the background and retaining the

original background: the temporal evolution of the

comparison between the current frame and the orig-

inal background is used to detect the event of interest.

A more sophisticated approach is proposed in (Tian

et al., 2011), which performs stopped object detection

through 4 steps: MOG-based background subtraction,

abandoned and removed object classiﬁcation, object

classiﬁcation (human, vehicle, package) and tracking.

In (Albiol et al., 2011) the authors propose a

parked vehicle detection algorithm, based on corner

detection, which creates a set of spatiotemporal maps,

used to understand what is happening in the scene and

to extract information such as the number of available

stalls, the number of parked vehicles in prohibited ar-

eas, the mean stopping times, the queue length and so

on.

The analysis of the literature performed up to now

makes it evident that the most successful methods are

those able to provide an accurate background sub-

traction. However, in order to achieve a good per-

formance, it is advisable to use a tracking algorithm,

possibly robust to occlusions, in order to evaluate the

movement of an object in the scene in a more sophisti-

cated way. Starting from these considerations, in this

paper we propose a hybrid solution that involves the

use of a novel background updating technique based

on a spatio-temporal analysis and of a tracking algo-

rithm based on objects similarity evaluation (Di Las-

cio et al., 2012)(Foggia et al., 2013). The background

image is modeled using adaptive selective updating,

whose adaptive weights depend on the time spent by

the particular pixel inside the scene. In particular,

we generate a grayscale heat map, where the inten-

sity of each pixel grows proportionally with its persis-

tence time in the foreground mask. The main idea is

that pixels corresponding to objects crossing the scene

have not to inﬂuence the background, and thus their

updating weight should be very low. On the contrary,

pixels belonging to objects stopped in the scene for a

long time should enter the background in a very fast

way, so their updating rate should be very high.

Furthermore, differently from traditional methods,

the tracking is performed on the heat map instead of

the foreground mask, so as to make the proposed ap-

proach insensitive to occlusion problems; indeed, any

persons who temporarily exclude a left object from

the camera view can not accumulate a sufﬁcient dwell

time to enter the heat map and thus is not tracked by

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

386

Figure 1: Overview of the proposed approach: once it has extracted the foreground mask and generated the heat map, the

system keeps track of each abandoned object and raises an alarm if its dwell time exceeds a threshold.

the system. In this way, the proposed algorithms is

able to detect abandoned objects by keeping track of

those having a stopping time greater than a ﬁxed time,

chosen by the human operator during the conﬁgura-

tion step.

2 THE PROPOSED METHOD

An overview of the proposed approach is shown in

Figure 1. The pixels corresponding to moving objects

are extracted (Foreground extraction) and are used to

update the heat map (Heat map generation), so that

the longer a pixel remains in the foreground mask,

the brighter it appears in the heat map. The heat map

and the foreground mask are used by the background

updating module, which adjusts foreground pixels up-

dating weights according to the corresponding inten-

sity on the heat map; thus, an object enters slowly the

background when its intensity is low but, gradually,

it enters faster if its persistence time increases. Note

that such strategy allows the system to easily discard

spurious objects due to the detection step as well as

persons that do not stop in the scene and that do not

need to be further analyzed. Finally, the system per-

forms on the heat map the detection and the tracking:

if an object is found and it stops for a long time in the

scene, then the system will raise an alarm.

2.1 Heat Map Generation

The objects moving in the scene at the current frame

t are encoded by the so called foreground mask (see

Figure 2(c)), obtained by traditional background sub-

traction algorithms.

Let be D

(x, y) the distance between the the cur-

rent image I

and the background updated up to the

previous frame B

t−1

in the generic pixel (x,y):

(x, y) = |I

(x, y) − B

t−1

(x, y)| (1)

The foreground mask can be computed as follows:

(x, y) =

(

1 i f D

(x, y) ≥ τ

f m

0 otherwise

(2)

being in our experiments τ

f m

set to 25.

Note that F

does not encode any information

about the temporal variation, that is in this kind of

applications a very important and not negligible fea-

ture. For this reason, we also introduce an heat map

, able to encode in a single image a temporal analy-

sis (see Figure 2(d)). In more details, the heat map is

a grayscale image whose generic pixel is updated by

a weighted moving average:

(x, y) = α · F

(x, y) + (1 − α) · H

t−1

(x, y), (3)

where α is the heat map updating weight, whose value

depends on the latency time chosen by the user dur-

ing the conﬁguration step. In this way, we are able

to obtain a kind of transparency overlay of F

on H

Starting from H

, its binarized version H

bin

is com-

puted:

bin

(x, y) =

(

1 i f H

(x, y) ≥ τ

0 otherwise

(4)

Thus, the system can start verifying if an object is

abandoned only once he has entered into H

bin

, so sig-

niﬁcantly reducing the computational effort required

by the successive detection and tracking steps.

AMethodforDetectingLongTermLeftBaggagebasedonHeatMap

387

(a) (b)

(e) (f)

Figure 2: An example of the different modules involved in

the stopped object detection: (a) current frame, (b) back-

ground, (c) foreground mask, (d) heat map, (e) binarized

heat map, (f) detected objects overlaid on the current frame.

Note that the event has been detected, so the object can start

entering the background.

2.2 Background Updating

One of the main contributions of the proposed ap-

proach pertains to the deﬁnition of a novel back-

ground subtraction algorithm. In fact, traditionally

the background is only updated by evaluating the

foreground mask at the current frame. On the other

hand, our aim is to control the entering time of the

objects moving in the scene depending on their mo-

tion, and then on their history. The key idea is that

the lower is the time spent in the scene by an object,

the higher is the time required for updating the back-

ground.In order to achieve this aim, we decided to se-

lectively update the background depending on both

the heat map and the foreground mask. In fact, the

weights for pixels corresponding to moving objects

are dynamically updated, depending on the time spent

by the object in that particular position.

In particular, let H

be the static heat map, that

is the heat map evaluated only on the pixels corre-

sponding to moving objects at the current frame and

computed as follows:

(x, y) = H

(x, y) · F

(x, y) (5)

Starting from H

, we can evaluate the background

as follows:

(x, y) =











· I

(x, y) + α

· B

t−1

i f H

(x, y) = 0

(x, y) · I

(x, y) + α

(x, y) · B

t−1

otherwise

(6)

being α

(x, y) = 1 − α

(x, y). α

and α

(x, y) rep-

resent the updating weights for the background and

for the foreground pixels, respectively. Note that α

is not ﬁxed, but instead it depends on the particular

pixel (x, y). In fact, it is related to the corresponding

intensity value on the heat map:

(x, y) = α

· H

(x, y) (7)

As for the α values, they strongly depend on the mini-

mum dwell time chosen by the human operator during

the conﬁguration step. In particular, we consider that

a stopped object has to enter the background as soon

as the related event is detected.

2.3 Event Evaluation

Differently from other state of the art approaches, the

detection of connected components and the tracking

of the objects is performed directly on the binarized

heat map. In particular, the tracking algorithm pro-

posed in (Di Lascio et al., 2013) has been used for

our experimentations. The main advantage deriving

from the choice lies in the fact that only those ob-

jects stopped for a long time are involved in the track-

ing process, so making the system particularly robust

with respect to occlusions as well as especially suited

for working in real and crowded environments.

Objects are ﬁnally analyzed and their permanence

time in the heat map is properly evaluated. Once an

event of interest has been detected, an alarm is raised

to the human operator.

3 EXPERIMENTAL RESULTS

The proposed approach has been tested over two

standard and widely adopted datasets, namely the

CAVIAR Dataset (CAVIAR, 2003) and the Imagery

Library for Intelligent Detection Systems Abandoned

Baggage Dataset (i LIDS, 2007).

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

388

(a) (b)

Figure 3: A few images extracted from the datasets used for

testing the proposed approach: CAVIAR (a,b) and i-LIDS

(c,d). On the left (a,c) the regions of interest are overlaid.

The former is composed by 26 videos acquired in

an indoor environment and showing different scenar-

ios: in particular, 5 videos contain a left bag event,

while the remaining 21 contain other situations typ-

ically occurring in real scenarios, such as a person

walking alone, people meeting with others, ﬁghting

and so on. In this way, the proposed approach can

be evaluated not only in terms of events correctly rec-

ognized, but also in terms of situations wrongly rec-

ognized as left bag events (false positive). The res-

olution is half-resolution PAL standard (384x288, 25

fps) and compressed using MPEG2; the total video

sequence duration is about 17 minutes.

The i-LIDS Abandoned Baggage dataset (here-

inafter i-LIDS) consists of 3 video with different dif-

ﬁculty levels, namely easy, medium and hard. It has

been used during i-LIDS bag and vehicle detection

challenge, hosted by AVSS 2007. Each video has a

resolution of 720x570 and has been acquired at 25

fps; the total length of the videos is about 10 minutes.

The scene captured by the camera shows a railway

where a person leaves a baggage unattended for 60

seconds. According to the competition rules, the de-

tection area can be divided into 3 zones (near, mid and

far), so as to set different parameters depending on the

distance from the baggage.

A few examples for both the datasets are shown

in Figure 3, while the obtained results are reported

in Tables 1 and 2 for CAVIAR and i-LIDS datasets,

respectively.

In particular, as for the CAVIAR dataset, the re-

sults for each typology of event is reported; the ta-

ble can be read as follows: TD indicates that a left

bag event has been correctly detected (true detected),

Table 1: Results obtained on the CAVIAR dataset, in terms

of True Detected (TD) and False Detected (FD).

CAVIAR Dataset

Video TD FD

Left Bag

Left Bag 1/1 0

Left Bag At Chair 1/1 0

Left Bag Behind Chair 0/1 0

Left Bag Picked Up 1/1 0

Left Box 1/1 0

Walk

Walk 1 - 1

Walk 2 - 0

Walk 3 - 0

Browse

Browse 1 - 0

Browse 2 - 2

Browse 3 - 0

Browse 4 - 0

Browse While Waiting 1 - 0

Browse While Waiting 2 - 0

Rest

Rest Fall On Floor - 2

Rest In Chair - 0

Rest Slump On Floor - 1

Rest Wiggle On Floor - 3

Meet

Meet Crowd - 0

Meet Split 3rd Guy - 0

Meet Walk Split - 0

Meet Walk Together 1 - 0

Fight

Fight Chase - 2

Fight One Man Down - 3

Fight Run Away 1 - 0

Fight Run Away 2 - 0

while FD indicates that something in the scene has

been wrongly detected as a left bag event (false de-

tected). For instance, the ﬁrst row can be read as

follows: in the Left Bag video, one event out of one

has been correctly recognized (so implying that there

are not missing events) and no false positives have

been detected. We can note that abandoned baggages

have been successfully detected in 4 videos out of

5 of the Left Bag dataset sequences, as also shown

in Figure 4; the missed event is due to the fact that

the bag is hidden behind the chair, so it is not pos-

sible with any traditional algorithm based on either

tracking or background subtraction methodologies to

discover this kind of event. We can also note that a

few false alarms have been detected, for instance in

the videos Fight One Man Down and Fight Chase. It

is mainly due to the fact that in such scenarios there

are people stopping on the wall for a very long time,

therefore the system raises an alarm since no informa-

tion about the typology of the objects is provided. It

is evident that the introduction of a classiﬁcation step,

able to distinguish, for instance, bags from persons,

may avoid the generation of these errors. Note that

classiﬁcation step is made possible in the proposed

approach by the tracking algorithm, which allows to

keep track of the objects and eventually of their re-

AMethodforDetectingLongTermLeftBaggagebasedonHeatMap

389

Table 2: Results obtained over the i-LIDS dataset and compared with state of the art approaches.

i-LIDS Abandoned Baggage

AB Easy AB Medium AB Hard

Method TD FD TD FD TD FD

Proposed method 1/1 0 1/1 0 1/1 1

(Maddalena and Petrosino, 2013) 1/1 0 1/1 0 1/1 1

(Evangelio et al., 2011) 1/1 0 1/1 5 1/1 6

(Pan et al., 2011) 1/1 0 1/1 0 1/1 0

(Tian et al., 2011) 1/1 0 1/1 0 1/1 1

(a) (b)

Figure 4: The baggage found in the CAVIAR dataset over

the videos (a) Left Bag, (b) Left Bag At Chair, (c) Left Bag

Picked Up and (d) Left Box.

lated properties, such as the class it belongs to.

As for the i-LIDS dataset, the obtained results are

shown in the ﬁrst row of Table 2. We can note that

the event of interest is recognized in all the videos

(Easy, Medium and Hard) and only one false positive

has been detected in the Hard videos. Figures 5(a-c)

show the detected event. It is worth noting that the

baggage in the Hard sequence has been correctly de-

tected (Figure 5(c)) although the occlusion occurring

in the previous frames. On the other hand, Figure 5(d)

shows the false positive detected by the system, due to

the stopped legs of the person sitting on the bench. As

said for the CAVIAR dataset, also in this case the in-

troduction of a classiﬁcation step would make it pos-

sible to distinguish bags from non bags objects. In

order to further conﬁrm the effectiveness of the pro-

posed approach, a proper comparison with state of the

art methods has been carried out. As shown in Table

2, the achieved results are comparable both in terms

of true detected and false detected. Not that the only

method which does not detect any events of interest is

proposed by (Pan et al., 2011): it is due to the fact that

it includes a post processing step in order to analyze

(a) (b)

Figure 5: The baggages identiﬁed in the i-LIDS dataset in

videos (a) AB Easy, (b) AB Medium and (c) AB Hard. In

(d) the false detected object is reported.

the detected events of interest and further reduce the

number of false positives arisen by the system.

4 CONCLUSIONS

In this paper we proposed a novel approach based on

an advanced background subtraction algorithm for de-

tecting stopped objects in crowded environments. The

experimentation has been conducted over two stan-

dard datasets, namely the CAVIAR and the i-LIDS

datasets, and the obtained results, compared with state

of the art approaches, conﬁrm that the system is able

to reliably detect the events also in presence of oc-

cluding objects, typically affecting this kind of algo-

rithms. Furthermore, the high recognition rate is not

paid in terms of false positives generated by the sys-

tem, that are still low even if compared with state of

the art approaches. Future works include the possibil-

ity to apply a classiﬁcation algorithm to the detected

objects stopped in the scene for a long time, so as to

distinguish between persons and inanimate objects;

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

390

this would make possible a further reduction of the

number of false positives detected by the system and

thus would improve the usability for a human opera-

tor.

ACKNOWLEDGEMENTS

This research has been partially supported by

A.I.Tech s.r.l. (http://www.aitech-solutions.eu).

REFERENCES

Acampora, G., Foggia, P., Saggese, A., and Vento, M.

(2012). Combining Neural Networks and Fuzzy Sys-

tems for Human Behavior Understanding. 2012 IEEE

Ninth International Conference on Advanced Video

and Signal-Based Surveillance, pages 88–93.

Albiol, A., Sanchis, L., and Mossi, J. M. (2011). Detection

of parked vehicles using spatiotemporal maps. Intel-

ligent Transportation Systems, IEEE Transactions on,

12(4):1277–1291.

Bevilacqua, A. and Vaccari, S. (2007). Real time detec-

tion of stopped vehicles in trafﬁc scenes. In Advanced

Video and Signal Based Surveillance, 2007. AVSS

2007. IEEE Conference on, pages 266–270. IEEE.

Bhargava, M., Chen, C.-C., Ryoo, M. S., and Aggarwal,

J. K. (2007). Detection of abandoned objects in

crowded environments. In IEEE AVSS, pages 271–

276. IEEE.

Boragno, S., Boghossian, B., Black, J., Makris, D., and Ve-

lastin, S. (2007). A dsp-based system for the detection

of vehicles parked in prohibited areas. In Advanced

Video and Signal Based Surveillance, 2007. AVSS

2007. IEEE Conference on, pages 260–265. IEEE.

CAVIAR (2003). Caviar test case scenarios. http://groups.

inf.ed.ac.uk/vision/CAVIAR/CAVIARDATA1/.

Di Lascio, R., Foggia, P., Percannella, G., Saggese, A., and

Vento, M. (2013). A real time algorithm for people

tracking using contextual reasoning. Computer Vision

and Image Understanding, 117(8):892–908.

Di Lascio, R., Foggia, P., Saggese, A., and Vento, M.

(2012). Tracking interacting objects in complex sit-

uations by using contextual reasoning. In VISAPP (2),

pages 104–113.

Evangelio, R. H., Patzold, M., and Sikora, T. (2011). A

system for automatic and interactive detection of static

objects. In Person-Oriented Vision (POV), 2011 IEEE

Workshop on, pages 27–32. IEEE.

Foggia, P., Percannella, G., Saggese, A., and Vento, M.

(2013). Real-time tracking of single people and

groups simultaneously by contextual graph-based rea-

soning dealing complex occlusions. In Performance

Evaluation of Tracking and Surveillance (PETS),

2013 IEEE International Workshop on, pages 29–36.

Guler, S., Silverstein, J. A., and Pushee, I. H. (2007). Sta-

tionary objects in multiple object tracking. In Ad-

vanced Video and Signal Based Surveillance, 2007.

AVSS 2007. IEEE Conference on, pages 248–253.

IEEE.

i LIDS (2007). Abandoned baggage dataset. ftp://motinas.

elec.qmul.ac.uk/pub/iLids/.

Maddalena, L. and Petrosino, A. (2013). Stopped object de-

tection by learning foreground model in videos. IEEE

transactions on neural networks and learning systems,

24(5):723–735.

Pan, J., Fan, Q., and Pankanti, S. (2011). Robust abandoned

object detection using region-level analysis. In Im-

age Processing (ICIP), 2011 18th IEEE International

Conference on, pages 3597–3600. IEEE.

Porikli, F., Ivanov, Y., and Haga, T. (2008). Robust

abandoned object detection using dual foregrounds.

EURASIP Journal on Advances in Signal Processing,

2008:30.

Singh, A., Sawan, S., Hanmandlu, M., Madasu, V. K., and

Lovell, B. C. (2009). An abandoned object detec-

tion system based on dual background segmentation.

In Advanced Video and Signal Based Surveillance,

2009. AVSS’09. Sixth IEEE International Conference

on, pages 352–357. IEEE.

Smitha, H. and Palanisamy, V. (2012). Detection of station-

ary foreground objects in region of interest from trafﬁc

video sequences. International Journal of Computer

Science Issues, 9(2):194–199.

Stauffer, C. and Grimson, W. E. L. (1999). Adaptive

background mixture models for real-time tracking.

In Computer Vision and Pattern Recognition, 1999.

IEEE Computer Society Conference on., volume 2.

IEEE.

Tian, Y., Feris, R. S., Liu, H., Hampapur, A., and Sun, M.-T.

(2011). Robust detection of abandoned and removed

objects in complex surveillance videos. Systems, Man,

and Cybernetics, Part C: Applications and Reviews,

IEEE Transactions on, 41(5):565–576.

Venetianer, P. L., Zhang, Z., Yin, W., and Lipton, A. J.

(2007). Stationary target detection using the ob-

jectvideo surveillance system. In Advanced Video and

Signal Based Surveillance, 2007. AVSS 2007. IEEE

Conference on, pages 242–247. IEEE.

AMethodforDetectingLongTermLeftBaggagebasedonHeatMap

391