Preview

Video Tracking: Theory and Practice

 

1 What is video tracking?

    1.1 Introduction

    1.2 The design of a tracker

        1.2.1 Challenges in video tracking

        1.2.2 Main components of a video tracker

    1.3 Problem formulation

        1.3.1 Single target tracking

        1.3.2 Multiple target tracking

        1.3.3 Definitions

    1.4 Interactive vs. automated tracking

    1.5 Summary

 

2 Applications

    2.1 Introduction

    2.2 Media production and augmented reality

    2.3 Medical applications and biological research

    2.4 Surveillance and business intelligence

    2.5 Robotics and unmanned vehicles

    2.6 Tele-collaboration and interactive gaming

    2.7 Art installations and performances

    2.8 Summary

 

3 Feature extraction

    3.1 Introduction

    3.2 From light to useful information

        3.2.1 Measuring light

        3.2.2 The appearance of targets

    3.3 Low-level features

        3.3.1 Colour

        3.3.2 Photometric colour invariants

        3.3.3 Gradient and derivatives

    3.4 Mid-level features

        3.4.1 Edges

        3.4.2 Interest points and interest regions

        3.4.3 Uniform regions

    3.5 High-level features

        3.5.1 Modelling the background

        3.5.2 Modelling the object

    3.6 Summary

 

4 Target representation

    4.1 Introduction

    4.2 Shape representation

        4.2.1 Basic models

        4.2.2 Articulated models

        4.2.3 Deformable models

    4.3 Appearance representation

        4.3.1 Template

        4.3.2 Histograms

        4.3.3 Coping with appearance changes

    4.4 Summary

 

5 Localisation

    5.1 Introduction

    5.2 Single-hypothesis methods

        5.2.1 Gradient-based trackers

        5.2.2 Bayes tracking and the Kalman filter

    5.3 Multi-hypothesis methods

        5.3.1 Grid sampling

        5.3.2 Particle filter

        5.3.3 Hybrid methods

    5.4 Summary

 

6 Fusion

    6.1 Introduction

    6.2 Fusion strategies

        6.2.1 Tracker-level fusion

        6.2.2 Measurement-level fusion

    6.3 Feature fusion in a Particle Filter

        6.3.1 Fusion of likelihoods

        6.3.2 Multi-feature resampling

        6.3.3 Feature reliability

        6.3.4 Temporal smoothing

        6.3.5 Example

    6.4 Summary

 

7 Multi-target management

    7.1 Introduction

    7.2 Measurement validation

    7.3 Data association

        7.3.1 Nearest Neighbour

        7.3.2 Graph matching

        7.3.3 Multiple Hypothesis Tracking

    7.4 Random Finite Sets for tracking

    7.5 Probabilistic Hypothesis Density filter

    7.6 The Particle PHD filter
        7.6.1 Dynamic and observation models

        7.6.2 Birth and clutter models

        7.6.3 Importance sampling

        7.6.4 Resampling

        7.6.5 Particle clustering

        7.6.6 Examples

    7.7 Summary

 

8 Context modelling

    8.1 Introduction

    8.2 Tracking with context modelling

        8.2.1 Contextual information

        8.2.2 Influence of the context on the filter

    8.3 Birth and clutter intensity estimation

        8.3.1 Birth density estimation

        8.3.2 Clutter density estimation

        8.3.3 Tracking with contextual feedback

    8.4 Summary

 

9 Performance evaluation

    9.1 Introduction

    9.2 Analytical vs. empirical methods

    9.3 Ground truth

    9.4 Evaluation scores

        9.4.1 Localisation scores

        9.4.2 Classification scores

    9.5 Comparing trackers

        9.5.1 Target life-span

        9.5.2 Statistical significance

        9.5.3 Repeatability

    9.6 Evaluation protocols

        9.6.1 Low-level evaluation protocols

        9.6.2 High-level evaluation protocols

    9.7 Datasets

        9.7.1 Surveillance

        9.7.2 Human-computer interaction

        9.7.3 Sport analysis

    9.8 Summary

 

Epilogue

 

Further reading

 

Appendix A: Comparative results

    A.1 Single vs. structural histogram

        A.1.1 Experimental setup

        A.1.2 Discussion

    A.2 Localisation algorithms

        A.2.1 Experimental setup

        A.2.2 Discussion

    A.3 Multi-feature fusion

        A.3.1 Experimental setup

        A.3.2 Reliability scores

        A.3.3 Adaptive vs. non-adaptive tracker

        A.3.4 Computational complexity

    A.4 PHD filter

        A.4.1 Experimental setup

        A.4.2 Discussion

        A.4.3 Failure modalities

        A.4.4 Computational cost

    A.5 Context modelling

        A.5.1 Experimental setup

        A.5.2 Discussion

 

About the authors

 

Notation

 

Acronyms

 

Index