STUDY ANALYSIS ON TRACKING MULTIPLE OBJECTS IN PRESENCE OF INTER OCCLUSION IN UNMANNED LC

Advanced Computational Intelligence: An International Journal (ACII), Vol.3, No.3, July 2016
DOI:10.5121/acii.2016.3305 35
STUDY ANALYSIS ON TRACKING MULTIPLE
OBJECTS IN PRESENCE OF INTER OCCLUSION IN
UNMANNED LC
A.RISVIN1
, S.V.ANANDHI2,
,P.SAHANA3
1,3
Department of Computer Science and Engineering, Dr.Sivanthi Aditanar College of
Engineering, Tiruchendur.
2
Department of Computer Science and Engineering, Dr.Sivanthi Aditanar College of
Engineering, Tiruchendur.
ABSTRACT
The object tracking algorithm is used to tracking multiple objects in a video streams. This paper provides
Mutual tracking algorithm which improve the estimation inaccuracy and the robustness of clutter
environment when it uses Kalman Filter. using this algorithm to avoid the problem of id switch in
continuing occlusions. First the algorithms apply the collision avoidance model to separate the nearby
trajectories. Suppose occurring inter occlusion the aggregate model splits into several parts and use only
visible parts perform tracking. The algorithm reinitializes the particles when the tracker is fully occluded.
The experimental results using unmanned level crossing (LC) exhibit the feasibility of our proposal. In
addition, comparison with Kalman filter trackers has also been performed.
KEYWORD
Mutual tracking, collision avoidance model, inter occlusion
1. INTRODUCTION
1.1 RELATED WORK
Now a day’s automatic tracking becomes progressively more important in low –cost and small-
scale image sensors which are used to deploy surveillance system across unmanned Lc which
renders any manual check of videos a very extensive task[1]. The tracking failure causes by
occlusion, handling of background clutter environment and illumination changes. The search
mechanisms to locate the target can be classified into gradient descent method where some
prespecified motion models as Kalman filter. The main role in estimation process in both cases is
to match the appearance model of the region predicted by spatiotemporal model. Using Kalman
filter [3]in multiple object tracking does not predicting the overlap condition and also is to
estimate the state of a linear system where the state is implicit to the distributed by a Gaussian
method. Such trackers does not need a beyond a initial bounding box(pretraining) and frequently
yield locally optimal solutions in clear scenario of overlapping, shape deformations and lighting
conditions.

36
Nevertheless such filters [4] are also vulnerable to local minima which can cause divergence. In
addition that the appearance model of the target which might also change over time due to
changing in light condition. To overcome this problem using mutual tracking algorithm for
tracking multiple objects which is the simplest way is likely to use multiple independent tracker
(M.I.T), where each target is related to an independent tracker not considering of the state of the
other targets. Such approach has been working with some success to follow hockey players and
tracking multiple people in unmanned LC in real time[5].
2. MUTUAL OBJECT TRACKING
The mutual object tracker to track multiple objects when frequent interaction between such
objects occlusion problem and also the problem of identity switch. To overcome this problem
using mechanism for monitoring all pairs of target estimation [9]. Initially in mutual object
tracking measure the distance between any pair wise target estimates is constantly monitored. In
this algorithm four cases can be reported.
The distances are large then use multiple independent tracking which indicating the
absence of occlusion and target identity switch.
If distance is smaller than predefined threshold but without causing of two bounding box
then use adaptive collision avoidance model will be applied
Overlapping between the two target estimation which exists then a partial occlusion based
reasoning will be applied and distinguish the occluded target from the non occluded
target.
If the distance indicates a full occlusion which means one bounding box is fully included
into the other one then the full occlusion reasoning is activated, where one waits for
reappearance of the target. In figure 1 illustrates the mutual object tracking as below
Short Long
No yes
Figure 1. Flow chart of mutual tracking algorithm
2.1 ADAPTIVE COLLISION AVOIDANCE MODEL
The adaptive collision avoidance model estimates the distance of two targets is less than some
predefined threshold and the regions are non overlapping, then compute the weights of the
particles in the next frame will be sophisticated by consider both the distance to the other target
and dissimilarity of the appearance of the two targets[6].
Check
distance
Form pair wise M.I.T (normal)
Overlap
Trajectory
monitorin
g
Mutual Object
Tracking

37
Let as consider the two targets X1 and X2 objects in the video that one wants to track, with
particles {L(i)
X1(k) , w(i)
X1(k) .The weight of the current w(j)
X1 of the jth particle will be calculated as
w’ (j)
X1 = w(j)
X1 exp(λX1X2 .d(RX1
j,RX2
)) (2.1)
Where RX1
j and RX2
are bounding box regions associated to jth particle of target X1 and global
estimate of target X2 respectively. d(RX1
j,RX2
) is the distance
between the two bounding box. Similar
responding applies to particle j of target X2, where the counterpart of (2.1) is
w’ (j)
X2 = w(j)
X2 exp(λX1X2 .d(RX2
j,RX1
)) (2.2)
Where
λX1X2 –similarity of the appearance models associated to the two targets.
2.2 HIERARCHICAL TRACKING PRIORITY
In the case partial occlusion the minor overlapping this cannot be excluded that the actual shapes
of two targets are completely identified without any uncertainty. If the two targets have similar
appearance model then use alternative reasoning will be required.
B (AX1
,AX2
)≥Ta (2.3)
The equation (2.3) indicates that the appearance models between the two target is deemed
To be similar and their similarity values in the sense of Bhattacharya coefficient is larger than
some predefined threshold Ta.
2.2.1 DIFFERENT TARGET APPEARANCE MODELS
Once the occlusion is conformed between the two bounding box regions then propose a two step
mechanism to conform the existence of the occlusion [8]. First compute the likelihood wx1 and
wx2 using the appearance model Ax1
and Ax2
around the global estimated bounding box of each
target. Second, to differentiate between the cases where the discrepancy of the two likelihood is
due to a wrong estimation.
Let us assume without loss of generality that,
wx1<wx2 (2.4)
Then calculate the positive contribution of the appearance model of another target X2 over the X1
reference model, which can be calculated by
h1
X1
= ∑ (Ax2(i)
-Arefx1
(i)) (2.5)
i,[A
x2
(i)-A
refX1
(i)]>0
Next repeat the same reasoning with enlarged region of target x1 and focusing on background
region, yielding
H2
X1
= ∑ (Ax2(i)
-Arefx1
(i)) (2.6)
i,[A
xb1
(i)-A
refX1
(i)]>0

38
where,AX1
b indicates the histogram associated to background region of the enlarged bounding box
pertaining to target x1.Comparing (2.5) and (2.6)
hx1
1 > hx1
2 (2.7)
In this X1 is assumed to be occluded by X2.otherwise the occlusion cannot be confirmed in the
current frame. If the occlusion is confirmed first stop enabling the adaptive collision avoidance
model and second to determine the visible and non visible parts in bounding box regions of
targets and third to recomputes the weights accordingly.
Target X1
Target X2
Figure 2: Overlapping of the two target estimates
In figure 2, a simple case of target overlapping which conforms a partial occlusion. This approach
which create a subdivision of each bounding box region to create for visible and non visible parts.
The initial outcome of such occlusion reasoning is twofold.
1)In the case of figure 2,ignoring any occlusion scenario the visible part of target X2 corresponds
to partitions 1,,4,while that of target X1corresponds to partition 2 and 4.This is called as visual
partition model .If the occlusion reasoning occluded that only target X2 is partially occluded ,then
the target X1 becomes fully visible ,while partition of 1,3,and 4 are the visible parts of target X2
2) The new weight is computed for each visible parts only. More formally each subdivision j of
the bounding box associated to target Xi. Next considering a set V of subdivision that belongs to
visible part, the new weight is computed for the target Xi as
w vXi = 1/|V| ∑ w(sj)
Xi (2.8)
j€V
With
w(sj)
Xi = eλaB
(AXi
sj,Aref
Xi) (2.9)
Where, visible in (2.8)
2.2.2 SAME TARGET APPEARANCE MODELS
The targets have a same appearance models then the above reasoning cannot be employed to
conform of an occlusion. In this method monitor the trajectory of the targets and adjusting the
weight of the particles according to the direction of the target movement and also considering the
size of the bounding box regions. First determine the direction of movement the target for this
use. First monitoring the velocity of the center of the bounding box region within a predefined
time window. Let (Vx,Vy)be the average velocity of a given tracker which is computed from
1
2
3 4
1 2
3
4
Overlap region

39
previous m frames then one can use the sign of the largest absolute values between Vx and Vy to
decide on the direction as it shown in table I.
Table I Direction Map.
Max(|Vx|,|Vy|) Sign of Vx,Vy Output(Directions)
Vx + Right
Vx - Left
Vy + Down
Vy - Up
2.2.3 FULL OCCLUSION BEFORE TARGET RETRACKING
In case of full occlusion relies on the concept of object performance, which suggests that a fully
occluded target will reemerge from its occluder[10].In addition it is typically known that the
weight associated to occluding target particles are usually low. The idea [10] is to randomly
reinitialize the particles of the occluded target around the occluder, from that the tracker can
easily capture the reappeared target immediately after its reappearance in Figure 3.
Figure 3: Full occlusion (small circles: reinitialize particles)
The tracker will compare the appearance of newly estimated tracker to the reference model for
confirming the reemerge tracker. The threshold value set Tr =0.8, then the target conform the
reappeared target. Otherwise the tracker will perform reinitialize the particles based upon the
occluded target position.
3. EXPERIMENTS AND RESULTS
In this section, we first test the unmanned Lc video for object identification using Kalman filter.
Using Kalman filter which does not predict the object overlapping shown in Figure 4
(a) Frame 12 (b) Frame 53
Tracker X1
Tracker X2

40
(c) Frame 12 (d)Frame 52
(e)Frame 53 (f) Frame 54
Figure 4: Multiple object tracking performance using Kalman filter.
In this frame 53 the occlusion is occurred but it does not overlapping the id switch problem. This
is major drawback of using Kalman filter in multiple object tracking.
To better multiple objects tracking in inter occlusion using mutual tacking algorithm for our
experiment. Figure 5 shows the performance of Mutual tracking algorithm on Unmanned Lc.
(a)Frame12 (b) Frame52
(a)Frame53 (c) Frame 54
Figure 5: Multiple object tracking performance using Mutual tracking Algorithm.

41
4. CONCLUSION
This study presents tracking multiple objects using mutual tracking algorithm.First,to deal with
arising uncertainty from background clutter and changing illumination of light in background, the
contextual information is considered by enlarging the boundary of the estimated target region and
comparing with current observation.Second,the particles distribution is considered through the
recursive estimation that restricts the effect of outliers on global estimate of the target. Third to
track multiple objects, although the intuitive use of multiple independent tracker M.I.T where
each tracker is associated to an individual target, cautious is required to avoid the problem of
occlusion or identity switch. In this deal with the problem by monitoring the distance between
trackers. In case of partial occlusion where each particle is partitioned into equal partition and
only the visible parts of the partition are used for tracking. In case of full occlusion, to reinitialize
the particles around the occluder to capture the reappearance of target. An experimental result
shows comparisons with Kalman filter and mutual tracking algorithm using Unmanned LC video
which deals with occlusions, clutter and the illumination change.
REFERENCES
[1] X. Wang, “Intelligent multi-camera video surveillance: A review,”Pattern Recognit. Lett., vol. 34,
no. 1, pp. 3–19, Jan. 2013.
[2] D. Comaniciu, V. Ramesh, and P. Meer, “Real-time tracking of non-rigid objects using mean shift,”
in Proc. IEEE Conf. CVPR, vol. 2. Jun. 2000,pp. 142–149.
[3] S.-K. Weng, C.-M. Kuo and S.-K. Tu, “Video object tracking using adaptive Kalman filter,” J. Vis.
Commun. Image Represent. vol. 17, no. 6,pp. 1190–1208, Dec. 2006.
[4] Y. Wu, J. Lim, and M.-H. Yang, “Online object tracking: A benchmark,”in Proc. IEEE CVPR, Jun.
2013, pp. 2411–2418.
[5] K. Nummiaro, E. Koller-Meier, and L. Van Gool, “An adaptive colorbased particle filter,” Image
Vis. Comput., vol. 21, no. 1, pp. 99–110,Jan. 2003.
[6] V. Papadourakis and A. Argyros, “Multiple objects tracking in the presence of long-term
occlusions,” Comput. Vis. Image Understand., vol. 114, no. 7, pp. 835–846, Jul.2010.
[7] K. Okuma, A. Taleghani, N. de Freitas, J. J. Little, and D. G. Lowe, “A boosted particle filter:
Multitarget detection and tracking,” in Proc.ECCV, 2004, pp. 28–39.
[8] W. Du and J. Piater, “Multi-camera people tracking by collaborative particle filters and principal
axis-based integration,” in Proc. ACCV, 2007, pp. 365–374.
[9] H. Ben Shitrit, J. Berclaz, F. Fleuret, and P. Fua, “Tracking multiple people under global appearance
constraints,” in Proc. IEEE ICCV, Nov. 2011, pp. 137–144.
[10] V. Papadourakis and A. Argyros, “Multiple objects tracking in the presence of long-term
occlusions,” Comput. Vis. Image Understand., vol. 114, no. 7, pp. 835–846, Jul. 2010.
[11] S. Pellegrini, A. Ess, K. Schindler, and L. Van Gool, “You’ll never walk alone: Modelling social
behavior for multi-target tracking,” in Proc. IEEE ICCV, Sep./Oct. 2009, pp. 261–268.
[12] S. Hare, A. Saffari, and P. H. S. Torr, “Struck: Structured output tracking with kernels,” in Proc.
ICCV, Nov. 2011, pp. 263–270.
[13] X. Mei and H. Ling, “Robust visual tracking using _1 minimization, “in Proc. IEEE ICCV, Sep./Oct.
2009, pp. 1436–1443.
[14] J. F. Henriques, R. Caseiro, P. Martins, and J. Batista, “Exploiting the circulant structure of tracking-
by-detection with kernels,” in Proc.ECCV, 2012, pp. 702–715.
[15] F. Pernici and A. Del Bimbo, “Object tracking by oversampling local features,” IEEE Trans. Pattern
Anal. Mach. Intell., vol. 36, no. 12, pp. 2538–2551, Dec. 2014.
[16] Video Benchmark Data Set. [Online]. Available:
http://cvlab.hanyang.ac.kr/tracker_benchmark/seq/Subway.zip, accessed Feb. 2015.

STUDY ANALYSIS ON TRACKING MULTIPLE OBJECTS IN PRESENCE OF INTER OCCLUSION IN UNMANNED LC

Recommended

More Related Content

What's hot (18)

Viewers also liked (15)

Similar to STUDY ANALYSIS ON TRACKING MULTIPLE OBJECTS IN PRESENCE OF INTER OCCLUSION IN UNMANNED LC (20)

Recently uploaded (20)

STUDY ANALYSIS ON TRACKING MULTIPLE OBJECTS IN PRESENCE OF INTER OCCLUSION IN UNMANNED LC