A Presentation
on
Object Tracking
Presented by
Shweta Kanhere-Banait
shweta.kanhere@gmail.com
Outline
• Introduction to Object tracking
• Definition
• Flowchart
• Tracking classifications
• Object Tracking Algorithms
• Comparison
• Challenges of object tracking
• Observations/Conclusion
• Application
• Future work
• References
• Thank you
Introduction to Object tracking
• Computer vision has received great attention over the last two decades.
• This research field is important not only in security related software, but also
in advanced interface between people and computers, advanced control
methods and many other areas.
• Object tracking play a key role .
Definition
Object tracking is a discipline within computer vision, which aims to track
objects as they move across a series of video frames. Objects are often people,
but may also be animals, vehicles or other objects of interest, such as the ball in
a game of soccer.
Flowchart
Tracking classifications
• Primitive geometric shapes
• Articulated shape models
• Skeletal models
• Point Tracking
• Kernel Tracking
• Silhouette Tracking
Point Tracking
Deterministic
Method
Maximum
velocity
Smooth motion
Proximal uniformity
Common motion
Proximity
Rigidity
Statistical or Probabilistic Method
Multiple object Tracking.
Single object
Tracking.
Kalman Filter
Particle Filter
Joint Probability Data
Association Filter (JPDAF)
Multiple Hypothesis Tracking
Kernel Tracking
Support Vector Machines
(SVM)
Template Tracking Mean Shift
Method
Layering Based- Matching
Silhouette-Based Object Tracking
• When object is represented by the outlines with only single solid color
in between the outlines made up of edges is known as silhouette.
• It is used where object cannot be represented by the simple geometric
shapes or by set of points.
• Silhouette is feature-less, therefore object model is created with help
of contour, edge information, color histogram etc.
Object Tracking Algorithms
 1. Absolute Differences 2. Census Method 3. Feature Based Method
• Mean-shift
• KLTP
• Condensation
• Tracking-Learning-Detection (TLD)
• Tracking Based on Boundary of the Object
Popular object tracking algorithms that use deep learning methods:
• SORT
• GOTURN
• MDNet
SORT
• Object detection Engine
• The algorithm tracks multiple objects in real time, associating the objects
in each frame with those detected in previous frames using simple
heuristics
…SORT
…SORT
Poor Good Excellent
…SORT
GOTURN
• Generic Object Tracking Using Regression Network
• GOTURN is trained by comparing pairs of cropped frames from
thousands of video sequences
…GOTURN
…GOTURN
• https://youtu.be/kMhwXnLgT_I
….MDNet
• Multi Domain Network (MDNet) is a CNN architecture that won the VOT2015
challenge.
• The objective of MDNet is to speed up training in order to provide real-time results
• https://youtu.be/zYM7G5qd090
Challenges of object tracking
• Re-identification
• Appearance and disappearance
• Occlusion (snow ,storm, snow on the ground, fog, air turbulence etc)
• Illumination
• Co-ordinates matching in case of multiple camera systems
• Pose variation of the object
• Motion blur
To perform tracking with these challenges in real time make tracking tedious
Observations
• Tracking approaches that employ a stable model can only accommodate
small changes in the object appearance but do not explicitly handle
severe occlusions or continuous appearance changes.
• A potential approach to overcome the limitation is to learn different
views of the object and later use them during tracking.
• A tracker that takes advantage of contextual information to incorporate
general constraints on shape and motion of objects will usually perform
better than the one that does not exploit this information.
• The capability to learn object models online may greatly increase the
applicability of a tracker
Application
• Video surveillance,
• Vision-based control
• Video compression
• Human computer interfaces
• Robotics
Future work
• The accuracy of object tracking could potentially increase by developing
methods for a more automatic selection process of features.
• We know from experience that a human tends do make more mistake than a
computer program optimized for a certain purpose.
• Automatic feature selection has received attention in the area of pattern
recognition, where methods for this purpose are divided into filter methods
and wrapper methods.
• However, these have not gotten the same attention in the area of object
tracking, where feature selection still is mostly done manually.
• There could be room for improvement in object tracking by developing fast
and accurate methods for automatic feature selection.
References:
• Z. Kalal, K. Mikolajczyk, and J. Matas, “Tracking-Learning-Detection,” Pattern Analysis and Machine Intelligence, 2011.
• Moving Object Detection Approaches, Challenges and Object Tracking
• https://missinglink.ai/guides/computer-vision/object-tracking-deep-learning/
• Li, F. Li, and M. Zhang, “A Real-time Detecting and Tracking Method for Moving Objects Basedon Color Video,” in 2009 Sixth
International Conference on Computer Graphics, Imaging andVisualization. IEEE, 2009, pp. 317–322.
• Object Detection and Tracking, Fatih Porikli and Alper Yilmaz
• Abdurrahman, "Smart video-based surveillance: Opportunities and challenges from image processing perspectives," 2016 3rd
International Conference on Information Technology, Computer, and Electrical Engineering (ICITACEE), Semarang, 2016, pp.
10-10
• W. Kim and C. Jung, "Illumination-Invariant Background Subtraction: Comparative Review, Models, and Prospects," in IEEE
Access, vol. 5, pp. 8369-8384, 2017.
• Aseema Mohanty and Sanjivani Shantaiya. Article: A Survey on Moving Object Detection using Background Subtraction Methods
in Video. IJCA Proceedings on National Conference on Knowledge , Innovation in Technology and Engineering
(NCKITE 2015)NCKITE, 2015(2):5-10, July 2015
• M. Zhu and H. Wang, "Fast detection of moving object based on improved frame-difference method," 2017 6th International
Conference on Computer Science and Network Technology (ICCSNT), Dalian, 2017, pp. 299-303.
• Nesne Takibinde Uyarlanabilir Arama Alanı Adaptive Search Area in the Object Tracking , Kazım HANBAY , Bingöl
Üniversitesi, Bilgisayar Mühendisliği Bölümü, Bingöl, Türkiye , Bingöl, Türkiye
Thank you!
Discussion !

Object tracking final

  • 1.
    A Presentation on Object Tracking Presentedby Shweta Kanhere-Banait shweta.kanhere@gmail.com
  • 2.
    Outline • Introduction toObject tracking • Definition • Flowchart • Tracking classifications • Object Tracking Algorithms • Comparison • Challenges of object tracking • Observations/Conclusion • Application • Future work • References • Thank you
  • 3.
    Introduction to Objecttracking • Computer vision has received great attention over the last two decades. • This research field is important not only in security related software, but also in advanced interface between people and computers, advanced control methods and many other areas. • Object tracking play a key role .
  • 4.
    Definition Object tracking isa discipline within computer vision, which aims to track objects as they move across a series of video frames. Objects are often people, but may also be animals, vehicles or other objects of interest, such as the ball in a game of soccer.
  • 5.
  • 6.
    Tracking classifications • Primitivegeometric shapes • Articulated shape models • Skeletal models • Point Tracking • Kernel Tracking • Silhouette Tracking
  • 7.
    Point Tracking Deterministic Method Maximum velocity Smooth motion Proximaluniformity Common motion Proximity Rigidity Statistical or Probabilistic Method Multiple object Tracking. Single object Tracking. Kalman Filter Particle Filter Joint Probability Data Association Filter (JPDAF) Multiple Hypothesis Tracking
  • 8.
    Kernel Tracking Support VectorMachines (SVM) Template Tracking Mean Shift Method Layering Based- Matching
  • 9.
    Silhouette-Based Object Tracking •When object is represented by the outlines with only single solid color in between the outlines made up of edges is known as silhouette. • It is used where object cannot be represented by the simple geometric shapes or by set of points. • Silhouette is feature-less, therefore object model is created with help of contour, edge information, color histogram etc.
  • 10.
    Object Tracking Algorithms 1. Absolute Differences 2. Census Method 3. Feature Based Method • Mean-shift • KLTP • Condensation • Tracking-Learning-Detection (TLD) • Tracking Based on Boundary of the Object Popular object tracking algorithms that use deep learning methods: • SORT • GOTURN • MDNet
  • 11.
    SORT • Object detectionEngine • The algorithm tracks multiple objects in real time, associating the objects in each frame with those detected in previous frames using simple heuristics
  • 12.
  • 13.
  • 14.
  • 15.
    GOTURN • Generic ObjectTracking Using Regression Network • GOTURN is trained by comparing pairs of cropped frames from thousands of video sequences
  • 16.
  • 17.
  • 18.
    ….MDNet • Multi DomainNetwork (MDNet) is a CNN architecture that won the VOT2015 challenge. • The objective of MDNet is to speed up training in order to provide real-time results • https://youtu.be/zYM7G5qd090
  • 21.
    Challenges of objecttracking • Re-identification • Appearance and disappearance • Occlusion (snow ,storm, snow on the ground, fog, air turbulence etc) • Illumination • Co-ordinates matching in case of multiple camera systems • Pose variation of the object • Motion blur To perform tracking with these challenges in real time make tracking tedious
  • 22.
    Observations • Tracking approachesthat employ a stable model can only accommodate small changes in the object appearance but do not explicitly handle severe occlusions or continuous appearance changes. • A potential approach to overcome the limitation is to learn different views of the object and later use them during tracking. • A tracker that takes advantage of contextual information to incorporate general constraints on shape and motion of objects will usually perform better than the one that does not exploit this information. • The capability to learn object models online may greatly increase the applicability of a tracker
  • 23.
    Application • Video surveillance, •Vision-based control • Video compression • Human computer interfaces • Robotics
  • 24.
    Future work • Theaccuracy of object tracking could potentially increase by developing methods for a more automatic selection process of features. • We know from experience that a human tends do make more mistake than a computer program optimized for a certain purpose. • Automatic feature selection has received attention in the area of pattern recognition, where methods for this purpose are divided into filter methods and wrapper methods. • However, these have not gotten the same attention in the area of object tracking, where feature selection still is mostly done manually. • There could be room for improvement in object tracking by developing fast and accurate methods for automatic feature selection.
  • 25.
    References: • Z. Kalal,K. Mikolajczyk, and J. Matas, “Tracking-Learning-Detection,” Pattern Analysis and Machine Intelligence, 2011. • Moving Object Detection Approaches, Challenges and Object Tracking • https://missinglink.ai/guides/computer-vision/object-tracking-deep-learning/ • Li, F. Li, and M. Zhang, “A Real-time Detecting and Tracking Method for Moving Objects Basedon Color Video,” in 2009 Sixth International Conference on Computer Graphics, Imaging andVisualization. IEEE, 2009, pp. 317–322. • Object Detection and Tracking, Fatih Porikli and Alper Yilmaz • Abdurrahman, "Smart video-based surveillance: Opportunities and challenges from image processing perspectives," 2016 3rd International Conference on Information Technology, Computer, and Electrical Engineering (ICITACEE), Semarang, 2016, pp. 10-10 • W. Kim and C. Jung, "Illumination-Invariant Background Subtraction: Comparative Review, Models, and Prospects," in IEEE Access, vol. 5, pp. 8369-8384, 2017. • Aseema Mohanty and Sanjivani Shantaiya. Article: A Survey on Moving Object Detection using Background Subtraction Methods in Video. IJCA Proceedings on National Conference on Knowledge , Innovation in Technology and Engineering (NCKITE 2015)NCKITE, 2015(2):5-10, July 2015 • M. Zhu and H. Wang, "Fast detection of moving object based on improved frame-difference method," 2017 6th International Conference on Computer Science and Network Technology (ICCSNT), Dalian, 2017, pp. 299-303. • Nesne Takibinde Uyarlanabilir Arama Alanı Adaptive Search Area in the Object Tracking , Kazım HANBAY , Bingöl Üniversitesi, Bilgisayar Mühendisliği Bölümü, Bingöl, Türkiye , Bingöl, Türkiye
  • 26.
  • 27.