Martin Halvorsen

1
Content-based Lecture Video Indexing

• Martin Halvorsen

2
Content

• Introduction
• Research Questions
• Background
• Proposed System
• Implementations
• Conclusions

3
Introduction
4
Introduction

• Blackboard presentations considered as essential
  and indispensable
• Hard to navigate through traditional videos
• Lack of expertise and time consuming

5
Research questions
6
Research questions

• Q1 How can foreground/background segmentation in
  a lecture video work for different writing-boards?

7
Research questions

• Q2 How to automatically extract meta-data from
  lecture videos?

8
Research questions

• Q3 How to use such meta-data for indexing and
  searching of lecture videos?

9
Background
10
Background

• Tracking, detection and separation algortihms
• Motion estimation, background models, statistical
  estimated models

11
Background

• Eirik Grythe segmentation and text detection
• Synne Repp simulated indexing alg.

12
Proposed System
13
Proposed system

• Superior blockdiagram of the proposed system

Blackboard extraction
Metadata Extraction
Index Generation
F/B Segmentation
14
Teacher segmentation
15
Teacher segmentation

• Motion estimation - SAD

16
Teacher segmentation

• Morphological operations to close holes in the
  detection

17
Teacher segmentation

• Using motion history to improve the segmentation
  algorithm

18
Teacher segmentation

• Replace blocks that has motion in it

19
Teacher segmentation

• Demonstration

20
Meta-data extraction
21
Meta-data extraction

• What is meta-data?
• How to extract blackboard content?
• Statistics

22
Meta-data extraction

• Demonstration

23
Indexing
24
Indexing

• When to extract an image?
• Search a lecture video using extracted features

25
Conclusions
26
Conclusions

• Adaptive segmentation possible by using motion
  estimation
• Extraction of content is possible using image
  difference
• Meta-data is possible to extract using a
  foreground model to compare against new
  content
• Statistics of extracted meta-data can be used to
  automatically index a lecture video

27

• The end