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

net/publication/278729405

Occlusion Detection and Handling: A Review

Article in International Journal of Computer Applications · June 2015

DOI: 10.5120/21264-3857

CITATIONS READS

48 8,402

2 authors:

Himanshu Chandel Sonia Vatta

Jaypee University of Information Technology BAHRA University
2 PUBLICATIONS 49 CITATIONS 4 PUBLICATIONS 72 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Himanshu Chandel on 14 May 2018.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Applications (0975 – 8887)
Volume 120 – No.10, June 2015

Occlusion Detection and Handling: A Review

Himanshu Chandel Sonia Vatta

Bahra University Bahra University
Waknghat, Shimla Waknaghat, Shimla

ABSTRACT the robust solution [35]. Little and Gillet put the technique of
Object tracking and detection is a classical research area in the suppression using ordering constraint [36]. Identifying the
field of computer vision from decades. Numerous kinds of occlusion regions, by defining the two occlusion maps, that is
applications are dependent on the area of object detection, occlusion map1 showing parts of image 1, that are not visible
such as advance driving assistance system, traffic in image 2 and similarly occlusion map 2 in image 2, while
surveillance, scene understanding, autonomous navigation etc. the displacement field between the images act as a key part in
Many challenges still exist while detecting an object such as this proficiency [37].
illusion, low visibility, cast shadows and most importantly Occlusion is the basic element that limits the information in
occlusions of object. Occlusions occur under two categories, an image. In vision based systems researchers commonly deal
firstly its, self‐occlusion which means that, from a certain with binocular or stereo vision. Binocular is the process of
viewpoint, one part of an object is occluded by another part. obtaining depth information from a pair of cameras. In stereo
Secondly, its inter-object occlusion which means when two vision occlusion occurs when a portion of the picture visible
objects being tracked occlude each other. We will review on one image is occluded in the other by the scene itself or, a
various occlusion handling methods that involved single and section of the scene near the image boundary moves out of the
multiple cameras according to their application. In short, the field of persuasion on the other picture. The core challenge in
objective of this paper is to deliberate in detail the problem of tracking is to accurately detect the object in a wide range of
occlusion in object tracking and provide a concise review for environments such as sunny and rainy conditions in driving
the problem of occlusion handling under different categories assistance systems. For autonomous navigation systems,
and identify new trends. whether its unmanned ground vehicles (UGV), unmanned
aerial vehicles (UAV) or unmanned underwater vehicles
General Terms (UUV), vision is playing a crucial role in path planning and
Computer vision, Driverless Cars, Algorithms, HOG, DPM. autonomous navigation. In comparison to range sensors such
as LIDAR and SONAR, camera is gaining attention for vision
Keywords based systems as cameras not only deliver geometry but also
Occlusion; tracking; scene; vision; unmanned; car. provide rich appearance information. One of the simplest
method for detecting occlusion is cross checking and
1. INTRODUCTION extrapolation [24]. Mehmood, Nawaz and Rao in addition to
In computer vision object detection is an active area of
the Bhattacharya Coefficient threshold used the normalized
research since decades. The interest is due to various
cross correlation method for handling occlusion. The same
applications such as advance driving assistance system [1],
technique was followed further, Normalized cross correlation
surveillance [2], scene understanding [3].The core challenge
is used at every step followed by subtracting the mean and
and the basic step in tracking is to accurately detect the object
dividing the standard deviation [17, 18]. Asymmetric
in different environments, but due to complex backgrounds,
information as well as geometric, photometric constraints and
weather conditions, cast shadows and occlusions it becomes
robust cost aggregation for occluded candidates have also
difficult to track an object. In this paper we will be targeting
been used for handling occlusions [19]. The main objective of
the problem of occlusion handling as it can't be avoided in
[20] is to detect and handle occlusion inside a tunnel and on
applications of object tracking. Many computer vision
real road using video as an input. In this paper Position of
algorithms suffer due to the presence of occluded objects in a
trajectories is estimated using Relative descriptive histogram,
scene. The region, which is occluded though, depends on the
which is an extended approach of HOG [10]. Image
camera viewpoint. In some scenarios angle of the camera can
segmentation is the core step for various computer vision
define which part is occluded and which one is not, hence
algorithms and the same technique has been adopted by
minimization approach, temporal selection, graph cut method
Enzweiler, Eigenstetter, Schiele and Gavrila for handling
and sum of squared distance are followed for handling the
occlusion using Mean shift algorithm. Artificial neural
same problem of occlusion [15]. The very basic technique for
networks are trained on intensity depth and motion features.
predicting occlusion in tracking was started from the
The expert weights are computed that are related to the degree
geometric information [28, 29]. One solution to occlusion
of visibility of the associated component [21]. In stereo vision
problem is to use the fusion of multiple cameras to determine
occlusion is the classical research problem as images are
depth and further estimating the occluded part [30]. Depth
captured from different angles and some pixels are available
estimation itself is a challenging problem since local features
only in one image. The same approach is used for handling
alone are insufficient to estimate depth at a point and needs to
occlusion that is left and right disparity maps are obtained
consider the global context of an image, hence a polynocular
with the help of optimization based on modified constant
stereo algorithm is used for occlusion handling [33]. Rosales
space belief propagation [22, 23]. Scene understanding started
and Sclaroff presented the method of temporary and trajectory
with the dream of constructing machines that can see like
prediction for the same problem [31].Geometry of a scene
humans to understand general principles and current situations
plays a vital character in handling occlusion, the technique of
from images, but it has become lot more extensive than that.
epipolar lines and disparity map are used by Davi et.al to get

33
 International Journal of Computer Applications (0975 – 8887)
Volume 120 – No.10, June 2015

Applications such as image search engines, autonomous feature framework to improve the detection of occluded
driving, computational photography, vision for graphics, objects. In this work, the new classifier is trained on the
human machine interaction, were unexpected and other background segments by using cars with an occlusion up to
applications keep rising as scene understanding technology 80% [8]. In another work for handling structured occlusion in
develops. As a center setback of high level computer vision, traffic scenes, SIFT descriptors are computed for an input
while it has enjoyed some great success in the past 50 years, a image and holistic classifiers are used. These classifiers are
great deal more is needed to achieve a perfect understanding subdivided into three vertical parts (C_3 split) and each
of visual scenes. Christian Wojek concentrates on reliable classifier is trained with car views. Occlusion is handled in
multi-object tracking from a moving platform in challenging two stages with a rate of 0% and < 80 % and as an extension,
real-world scenes with object - object occlusion using SVM depth information for individual car parts helps the classifier
(Support vector machine) and DPM (Deformable Parts to reason about typical occlusion patterns [9]. Fei Wu et al.
Model) which is good in long term occlusion, basically on uses an access for handling occlusion inside a tunnel road and
human detection [25]. Next, the goal of Fehrman and on real roads. According to them, partial occlusion still
McGoughis to fully remove the occluding objects from the occurs, but the trajectory can handle this situation and tracks
image and look at what is behind them using the canny edge the vehicles well. Firstly, older trajectory is estimated and the
detector function and disparity map generation [27]. Kalman estimated position of tracked vehicle is marked with gray.
filtering is the classical approach used widely for tracking. After updating older trajectories, newly generated trajectory
Appearance models are used for handling occlusions in each estimates the new position using relative discriminative
track showing how the object appears in an image [32]. histogram of oriented gradients particle filter (RDHOGPF).
Niknejad, Kawano, Oishi, Mita uses two layers classifiers Magnitudes are set to zero in the gray region using
using Conditional Random Field (CRF) and Deformable Parts
Model (DPM) without any knowledge from previous frames M (u,v)= if F(u,v)∼= 1 (1)
[26]. Novel approach has been adopted by Zhang, Xu,
Beaugendre and Goto for the same problem of handling
occlusions in human tracking based on Kannade-Lucas
where u and v are the horizontal and vertical coordinates,
Tomasi (KLT) method, they have used PETS and CAVIAR
respectively, M (u,v) and F(u,v) are the respective magnitude
dataset.
and foreground values in the normalized 32×32 image [20].
We further classify the various methods and techniques used Pang, Lam and Yang propose a novel method for handling
for occlusion handling on the basis of their applications in occlusion, which is tested on monocular traffic image
scene understanding. The applications are divided into four sequences. They have adopted the texture based segmentation
sections in section 2, we explain about the techniques used for and deformable modeling method which solves most of the
car occlusion in realistic multiple traffic and street scenes occlusion cases involving two vehicles but fails in severe
followed by human occlusion in various scenes in section 3. occlusion [38]. Qing, Hoang and Jo deals with the problem of,
Then, in section 4, diverse real world occluded scenes are same color vehicle occlusion as shown in figure1, using
reported and in section 5, miscellaneous applications are Bhattacharya distance and local edge image projection, for the
discussed. Finally, conclusion will be presented in section 6. separation of each occluded vehicle. For measuring the color
similarity authors have used the bhattacharya distance
2. CAR OCCLUSION IN STREET between two color histograms of occluded vehicles [39].
SCENES
Due to increase of traffic on roads, intelligent traffic 3. PEDESTRIAN OCCLUSION IN
surveillance systems are being implemented in various DIVERSED SCENES
countries for highway monitoring and city road management Occlusion is a challenging problem for tracking people under
system. Traditionally, shadow detection techniques have been various conditions. Human detection is very important in
employed for removing shadows from the background and video surveillance and intelligent vehicles such as UGV's and
foreground, but Sadeghi & Fathy used it as a novel feature for UAV's. Due to inconsistent manifestation and extensive series
vehicles detection and occlusion handling. They have used of poses which humans can espouse, detection of humans is a
photometric characteristic of darkest pixels of strong shadows challenging task whether in a picture or a video. In a real
in traffic images for occlusion handling [3]. Multilevel environment, noteworthy amount of partial occlusion occurs
framework is adopted by Zhang, Jonathan Wu, Yang and as pedestrians move in the immediacy of other objects.
Wang for handling car occlusion in traffic; firstly occlusion is Published literature on tracking system development is mostly
detected by evaluating the compactness ratio and interior focused on tracking an object moving in indoor environments.
distance ratio of vehicles, and then the detected occlusion is The remarkable progress in human detection and long-term
handled by removing a cutting region of occluded vehicles. tracking has allowed to detect and track numerous people
On interframe level, occlusion is detected by performing concurrently in intricate scenes. Yet, systems are still
subtractive clustering on the motion vectors of vehicle. Next relentlessly challenged by partial and full occlusions, which
comes the tracking level, occluded vehicles are tracked by occur commonly in scenes of realistic complexity.
using a bidirectional occlusion reasoning mechanism [4]. The
Rosales and Sclaroff reported a technique based on current 3D
same problem of occlusion was handled with the help of
estimates of velocity and position to predict future locations
geometric occlusion patterns and occlusion mask. The results
of objects. If a collision is detected in objects, assumption is
are quite interesting as for heavily occluded vehicles that are
made that occlusion is likely to happen if the current
only 30 – 60% visible, 83.1% of the occlusions are correctly
estimation of error is given by Extended Kalman Filter (EKF).
predicted on inner city street scenes datasets with different
The occlusion detection routine predicts future locations of
levels of occlusion [11]. In latest approaches for handling car
objects, based on current estimates of 3D velocity and
occlusion, different classifiers are used with the help of
position. By using estimated bounding box parameters and the
foreground and background segmentation. Straw, Hasler and
binary maps for the objects being tracked, occlusion is
Wersing uses car detection technique based on an analytical
predicted, detected and handled by estimating positions and

34
 International Journal of Computer Applications (0975 – 8887)
Volume 120 – No.10, June 2015

velocities followed by applying these estimates to the image 4. DIVERSE REAL WORLD
plane [31].
OCCLUDED SCENE
Chang, Gong and Ong attempted to handle the occlusion with In stereo images, to deal with occlusion, Kang, Szeliski and
multiple cameras. In this paper, tracking condition is divided Chai uses two approaches. Firstly, they used adaptive
in two parts: one camera occlusion (OCO) and two camera windows and selected a temporal subset of the frames to
occlusion (TCO). In OCO, geometric and recognition based match at each pixel. For explicitly modeling occluded regions,
methods are used with the help of various cues such as Global (MRF) minimization based on graph cuts was used as
epipolar line, landmarks, apparent color and height [30]. the second approach [15]. Sun, Li, Kang and Shum handled
Condensation algorithm [30, 6] was used to track predicted the same problem of occlusion using an asymmetric stereo
joint features of the subject. Hu and Tan use the approach of matching model with visibility constraint. Two steps are
Bayesian network for solving occlusion in tracking of people. considered; first the disparity map was deduced taking into
The occlusion relation is categorized by ϕt∈ {0, 1, 2}, where consideration the occlusion map of another view. Then
ϕt=0 indicates no occlusion, ϕt=1indicates A occludes B, and concluding the occlusion map in one view from the disparity
ϕt=2 indicates B occludes A [30]. Wang, Han and Yan map of another view. Linear support vector machine and
developed a novel and impactful method for human detection segmentation by the mean-shift approach is also used in
and handling partial occlusion using the combination of handling occlusions [13]. Many of the algorithms are
Histogram of Gradient (HOG) and Local Binary Pattern designed to deal with short-duration partial occlusion, but
(LBP). For scanning, whole windows global detector is used they fail at severe occlusions and when a partial occlusion
and for local regions, part detectors are used. With the help of lasts for a long time. Yang, Li, Pan and Li used an approach
the augmented HOG-LBP feature and the global-part for multiple object tracking in dynamic scenes. The developed
occlusion handling method, they have achieved a detection technique which has the ability to cope with long-duration and
rate of 91.3% with FPPW= 10−6, 94.7% with FPPW= 10−5, complete occlusion without a prior knowledge about the shape
and 97.9% with FPPW= 10−4 on the INRIA and Pascal or motion of objects. Experiments are implemented on
dataset [10]. CAVIAR and PETS dataset in complex indoor and outdoor
environments. The states of objects are separated into three
Vision based autonomous navigation is becoming more and
parts: before, during and after occlusion. This Real time
more popular in unmanned vehicles, whether its air, ground or
system uses video as an input and performs object
water. There are various issues and difficulties in vision based
segmentation using merging and splitting technique [18].
navigation systems such as weather conditions, cast shadows,
strong headlights and complex backgrounds. One of the main Object tracking has various applications in computer vision.
setback in navigating the vehicle autonomously is occlusion Pan and Hu propose content adaptive progressive occlusion
caused by other vehicles and various objects. Ess, Schindler, analysis algorithm for object tracking with occlusion
Leibe and Gool have reported an approach for autonomous handling. The entire Region of interest is determined, then the
navigation and path planning, but most importantly, they have occlusion status of the block containing undetermined pixels
targeted on multiperson detection and targeting. Firstly, they are analyzed and finally all such blocks are scanned [5].
generate the occlusion map that contains the regions occluded Ramik, Sabourin, Moreno and Madani developed an
by both pedestrians and static obstacles. Secondly, Extended intelligent vision system that has the ability to learn
Kalman Filter (EKF) increases the chances of finding the autonomously individual objects present in a real environment
object once it becomes visible again. In figure 2, the boxes are without using human labeled data. Visual saliency is the
color coded to show tracked identities, darker boxes indicate nucleus of this paper for detection and recognition. SURF is
objects in occlusion [12]. used for object recognition as well as occlusion [14]. For
handling occlusion while tracking, Guan, Chen, Yang and Wu
Enzweiler, Eigenstetter, Schiele and Gavrila uses the expert
implemented the occlusion handling component in three
weights that are related to the degree of visibility for handling
stages: occlusion detection, person re identification and
partial occlusion. The main objective of the reported study is
tracklet generation. Before and after occlusion information is
pedestrian classification using images as input. The
used to model the association cost matrix to re-identify
component based expert classifiers are trained on intensity
persons under strong occlusion [17]. Chen, Li, D. Zhao and Q.
depth and motion features which form the core of the used
Zhao gives a remarkable model shown in figure 3 for handling
framework. Mean-shift segmentation with depth and motion is
occlusion, inspired by the human perception of occlusion,
applied online for recovering occlusion dependent weights
further adding adaboost algorithm. They have used five
[21].
dimensional region based cues; semantic, position,
Stereo vision is a good approach for handling occlusions but compactness, shared boundary and junction cue on rural
Wojek, Walk, Roth and Schiele, have reported a monocular artificial and outdoor scenes [40].
vision based scene understanding on the ETH-PedCross2
dataset. Multi-object tracking from a moving platform in 5. MISCELLEANOUS
challenging real-world scenes and object-object occlusion for The Weng, Ahuja and Huang reported an approach for
human detection in ETH-PedCross2 have been achieved using detecting occlusions in 1988 by defining two occlusion maps,
seven detector components, deformable parts model and whose working is shown in figure4. Black areas represent
support vector machine (SVM). The reported technique was occlusion regions. The displacement field from image 2 to
proved to be good for handling long term occlusions [25]. Hu image 1, without occlusion information is determined with the
and Tan uses the Bayesian Network approach for handling help of computed occlusion map1. Some of the areas in image
occlusions. Human model is compared with 2D ellipse model 1 that have not been matched are occluded in image 2 and are
and tracking process is done using the condensation algorithm mapped in occlusion map 1 [37]. When the location of the
[6]. occluded region depends on the viewpoint, then the match
suppression technique is used for occlusion handling [36].
Occlusions are the regions in one image that have no match in
the other image. Geiger, Ladendorf and Yuille use the

35
 International Journal of Computer Applications (0975 – 8887)
Volume 120 – No.10, June 2015

technique of stereo with an occlusion process and for the solution is not there to handle occlusions in every scene. Lot
of research has to be done to make it more accurate. The best
right such that: [35] feature we can conclude is shape as it helps to match occluded
part of the object with the same object in the next frame.
(M) = 1- (2)
Various methods under stereo vision are showing some good
progress in handling occlusions. Occlusion is a challenging
(M) = 1- (3)
problem in object detection and when it comes to stereo
matching, it can’t be neglected at all. Estimating disparity and
The further disparity field is defined as for the left eye
occlusion by using disparity and occlusion maps is the
and the other for right eye by: classical solution of this problem and till date we don’t have
100% accuracy in handling occlusions under various
(M) = 1- if = 0 (4) scenarios. Geometric patterns with the combination of
machine learning is playing an effective role for handling
(M) = 1- if = 0 (5) occlusions. HOG has gained attention for object detection but
Wunsch and Hirzinger used Kalman Filtering for estimating when it comes to handling occlusion results are still
object velocity and acceleration. Geometric information was unsatisfactory. Models inspired by human perception can lead
used for predicting occlusions [28]. LathuiIiere and Hervze to satisfactory occlusion handling results. In the end, we
handled the occlusion in application of hand posture. An would conclude it in two points, there is still a large gap
occlusion validation algorithm was proposed which consisted between human perception and occlusion reasoning. Lastly, if
of three parts: self-occlusion, sphere occlusion and cylinder the detected features are better, the occlusion estimation will
occlusion [29]. Egnal and Wildes used five different be better.
algorithms for the purpose of detecting occlusion. Bimodality
(BMD) and Match Goodness Jumps (MGJ) are the first two
algorithms that detect occlusion boundaries and the rest are
Left-Right Checking (LRC), Ordering (ORD) and occlusion
constraint (OCC) [24]. Min and Sohn used geometric and
photometric constraints for handling occlusion. Disparity map
was estimated before asymmetric occlusion detection and the
set occluded pixels were detected by the same technique [19].
From the dataset of Tsukuba, Venus, Teddy and Cone, images
have been used by Jang and Hoe for handling occlusion using
warping constraint and cross check constraint techniques. In
this algorithm, the authors indicated that the pixels with
largest disparity values are selected as visible pixels and rest
matching pixels become occluded [22,23].

6. CONCLUSION
This paper provides a summarizing study for handling
occlusions under different scenarios. We have tried to show
the various Image processing, Computer vision and Machine
Learning techniques for the same problem of occlusion. When
both the objects occluded and occluder are being tracked by Fig.1.Same color vehicle occlusion [39]
the same algorithm, it does not perform well in handling
occlusions. DPM alone fails to handle occlusion, many
methods have been used on various datasets, but still optimal

Fig.2.Pedestrian tracking [12]

36
 International Journal of Computer Applications (0975 – 8887)
Volume 120 – No.10, June 2015

Fig.3.Overview of the method [40]

[9] Marvin Struwe1 Stephan Hasler and Ute Bauer-Wersing,

7. ACKNOWLEDGMENTS “A Two-Stage Classifier Architecture for Detecting
Our thanks to the faculty and all the members of lab who have Objects under Real-WorldOcclusion Patterns,” Artificial
contributed towards development of this survey. Neural Networks and Machine Learning – ICANN 2014

8. REFERENCES [10] Xiaoyu Wang,Tony X. Han and Shuicheng Yan, “An

[1] Hossein Tehrani Niknejad, Taiki Kawano,Yuki Oishi, HOG-LBP Human Detector with Partial Occlusion
Seiichi Mita, “Occlusion Handling using Discriminative Handling,” IEEE 12th International Conference on
Model of Trained Part Templates and Conditional Computer Vision (ICCV),2009
Random Field,” 2013 IEEE Intelligent Vehicles [11] M. Zeeshan Zia, Michael Stark and Konrad Schindler,
Symposium (IV) June 23-26, 2013, Gold Coast, “Explicit Occlusion Modeling for 3D Object Class
Australia Representations,” IEEE Conference on Computer Vision
[2] Andry Maykol Pinto, Paulo G. Costa and A. Paulo and Pattern Recognition,2013
Moreira, “An Architecture for Visual Motion Perception [12] A. Ess, K. Schindler, B. Leibe, L. van Gool, “Improved
of a Surveillance-based Autonomous Robot,” Multi-Person Tracking with Active Occlusion Handling,”
2014IEEEInternational Conference on Autonomous Proceedings of the IEEE ICRA Workshop on People
Robot Systems and Competitions(ICARSC) May14- Detection and Tracking,2009
15,Espinho,Portugal
[13] Jian Sun, Yin Li, Sing Bing Kang, Heung-Yeung Shum,
[3] Andrej Karpathy Li Fei-Fei, “Andrej Karpathy Li Fei- “Symmetric Stereo Matching for Occlusion Handling,”
Fei,” 2014 Computer Vision and Pattern Recognition, 2005.
[4] Whei Zhang,Q.M Jonathan Wu,Xiaokang Yang, [14] Dominik Maximilián Ramík,Christophe Sabourin,Ramon
“Multilevel framework to detect and handle Occlusion,” Moreno,Kurosh Madani, “A machine learning based
IEEE Transaction on Intelligent Transport intelligent vision system for autonomous object detection
Systems,VOL.9,NO.1,2008 and recognition,” Applied Intelligence, Volume 40, Issue
[5] Jiyan Pan,Bo Hu, “Robust Occlusion handling in object 2, pp 358-375, March 2014
tracking,” IOP Sciences,Computer Vision and Pattern [15] Sing Bing Kang, Szeliski R. , Jinxiang Chai, “Handling
Recognition,2007 occlusions in dense multi-view stereo,” Computer Vision
[6] Min Hu, Weiming Hu, Tieniu Tan, “Tracking People and Pattern Recognition, 2001
through Occlusion,” International Conference on Pattern [16] Rashid Mehmood,Rab Nawaz,Naveed Iqbal Rao,
Recognition,2004 “Occlusion Handling in Meanshift Tracking using
[7] M. Zeeshan Zia, Michael Stark and Konrad Schindler, Adaptive Window Normalized Cross Correlation,” 11th
“Are Cars Just 3D Boxes? – Jointly Estimating the 3D International Bhurban Conference on Applied Sciences
Shape of Multiple Objects,” IEEE Conference on and Technology (IBCAST), 2014
Computer Vision and Pattern Recognition,2014 [17] Yaowen Guan, Xiaoou Chen, Deshun Yang, Yuqian Wu,
[8] Marvin Struwe,Stephan Haslerand Ute Bauer-Wersing, “Multiperson tracking by detection with local particle
“Using the Analytic Feature Framework for the filtering and global occlusion handling,” IEEE
Detection of Occluded Objects,” Artificial Neural International Conference on Multimedia and Expo
Networks and Machine Learning – ICANN 2013 (ICME), 2014

37
 International Journal of Computer Applications (0975 – 8887)
Volume 120 – No.10, June 2015

[18] Tao Yang, Stan Z.Li, Quan Pan, Jing Li, “Real-time Occlusion,” Proceedings of the IEEEInternational
Multiple Objects Tracking with Occlusion Handling in Conference on Robotics and Automation,1997
Dynamic Scenes,” IEEE Computer Society Conference
on Computer Vision and Pattern Recognition,2005 [29] Fabienne LathuiIiere and Jean-Yves Herve, “Visual
Tracking of Hand Posture with Occlusion
[19] Dongbo Min and Kwanghoon Sohn, “Cost Aggregation Handling,”15th International Conference on Pattern
and Occlusion Handling With WLS in Stereo Matching,” Recognition,2000
IEEE Transactions on image processing, VOL. 17, NO.
8,2008 [30] Ting - Hsun Chang,Shaogang Gong and Eng Jon Ong,
“Tracking multiple people under occlusion using
[20] Bing-Fei Wu et al. , “A Relative-Discriminative- multiple cameras,” BMVC,2000
Histogram-of-Oriented-Gradients-Based Particle Filter
Approach to Vehicle Occlusion Handling and Tracking,” [31] Ro´merRosales and StanSclaroff, “Improved Tracking of
IEEE Transactions on Industrial Electronics, VOL. 61, Multiple Humans with Trajectory Prediction and
NO. 8,2014 Occlusion Modeling,” IEEE Conf. on Computer Vision
and Pattern Recognition,1998
[21] Markus Enzweiler,Angela Eigenstetter,Bernt
Schiele,Dariu M. Gavrila, “Multi-Cue Pedestrian [32] Andrew Senior et al. , “Appearance models for occlusion
Classification With Partial Occlusion Handling,” IEEE handling,” Image and Vision Computing ,”VOL.24 ,2006
Conference on Computer Vision and Pattern [33] Yasuyuki SUGAYA and Yuichi OHTA, “Stereo by the
Recognition, 2010 integration of two algorithms with/without occlusion
[22] Woo-Seok Jang , Yo-Sung Ho, “Efficient depth map handling,”15th International Conference on Pattern
generation with occlusion handling for various camera Recognition,2000
arrays,” Signal, Image and Video Processing, Volume 8, [34] Chenyuan Zhang, Jiu Xu, Axel Beaugendre and Satoshi
Issue 2, pp 287-297,2014 Goto, “A KLT-Based Approach for Occlusion Handling
[23] Woo-Seok Jang and Yo-Sung Ho, “Efficient Disparity in Human Tracking,”Picture Coding Symposium,2012
Map Estimation Using Occlusion Handling for Various [35] Davi Geiger,Bruce,Ladendorf,Alan Yuille, “Occlusions
3D Multimedia Applications,” IEEE Transactions on and Binocular Stereo,”International Journal of Computer
Consumer Electronics, 2011 Vision,1995
[24] Geoffrey Egnal and Richard P. Wildes, “Detecting [36] James J. Little and Walter E. Gillet, “Direct Evidence
Binocular Half-Occlusions:Empirical Comparisons of for Occlusion in Stereo and Motion,” Computer Vision
Five Approaches,”IEEE Transactions on pattern analysis — ECCV, 1990
and machine intelligence, VOL. 24, NO. 8, 2002
[37] Juyang Weng,Narendra Ahuja,Thomas S. Huang, “Two
[25] Christian Wojek, Stefan Walk, Stefan Roth and Bernt view matching,” IEEE,1988
Schiele, “Monocular 3D Scene Understanding with
Explicit Occlusion Reasoning,” IEEE Conference on [38] Clement Chun Cheong Pang, William Wai Leung Lam,
Computer Vision and Pattern Recognition (CVPR), 2011 and Nelson Hon Ching Yung,“A Novel Method for
Resolving Vehicle Occlusion in a Monocular Traffic-
[26] Hossein Tehrani Niknejad, Taiki Kawano,Yuki Oishi, Image Sequence,”IEEE Transactions on intelligent
Seiichi Mita, “Occlusion Handling using Discriminative transportation systems, VOL. 5, NO. 3,2004
Model of Trained Part Templates and Conditional
Random Field,” IEEE Intelligent Vehicles Symposium [39] Ming Qing, Van-Dung Hoang and Kang-Hyun Jo
,2013 Yung,“Localization and Tracking of Same Color Vehicle
under Occlusion Proble,” Mechatronics-REM, 2012
[27] Brian Fehrman and Jeff McGough,“Handling Occlusion
with an Inexpensive Array ofCameras,”IEEE Southwest [40] Xiaowu Chen, Qing li, Dongyue Zhao, Qinping
Symposium on Image Analysis and Interpretation Zhao,“Occlusion cues for image scene
(SSIAI),2014 layering,”Computer Vision and Image Understanding
Volume 117, Issue 1, Pages 42–55, 2013
[28] P. Wunsch and G. Hirzinger, “Real-Time Visual
Tracking of 3-D Objects with Dynamic Handling of

IJCATM : www.ijcaonline.org 38

View publication stats