臺灣博碩士論文加值系統

Object tracking is one of the major problems in computer vision. Tradition-
ally, tracking approaches can be roughly divided into two categories. Given
the initial object’s location in the first frame of sequential image or video, the
first approach tracks object based on studying the correlations between the ob-
ject’s information in consecutive frames. On the opposite, given some object’s
prior knowledge, the second approach detects the object in each single frame,
so called the “tracking by detection” approach. Both approaches suffer some
weakness while holding some advantages on the other side. The first approach
do not need prior knowledge about the object that will be tracked, however, it
generally fails when the object’s appearance is changing due to view changing,
occlusion or varying in scale when moving along the viewing axis. Based on a
built object model, the tracking-by-detection approach needs no information
from previous frames; therefore, it can deal with appearance changing if the
object model recognizes the appearance. However, the approach needs a ro-
bust model for the object but sometimes it is hard to be obtained. In general,
no matter what approach we try, we lack of “positive” samples, either samples
from previous frames in the first approach, or samples from the prior database.
In this work, we propose a tracker based on transfer learning which can take
advantage of knowing some similar, but not exactly identical objects’ informa-
tion; based on the mechanism, we combine information related to the tracked
object, and the information from similar objects to form a robust model for
object tracking.
We consider the tracking as a binary classification problem. The positive data is obtained from object’s appearance in previous frames, and from some
similar objects’ appearance in a database; the negative data is obtained from
background area in the first frames. All data is combined in a transfer learning
framework, and we adopt TrAdaBoost algorithm for our experiments. By the
proposed model, we can distinguish among several situations in tracking: a)
the tracker can follow the object; b) the tracker can not follow the object due
to object’s changing appearance; c) the tracker can not follow the object due
to occlusion. In the tracking process, we continuously compute the “fitting
score” between the model and the observed appearance. When the score is
high, we continue the tracking without modifying the model; when the score is
low, the time when we may easily loose the object, we need to separate between
two cases: the object changes it’s appearance (where we need to modify our
model) or the object is occluded by other objects (where we do not need to
modify our model). Thanks to the help of similar objects’ appearance from
the database, we can choose between a changing appearance or an occasional
appearance due to occlusion. Experiments show that the proposed method
works well. Given 150 frames testing data, this proposed method successfully
tracks the object in 90% of those frames, compared to 50% success rate for the
model that does not consider appearance changing. Moreover, it successfully
distinguishes between the appearance changing and occlusion and therefore
can greatly enhance the tracking performance.

1 Introduction 2
1.1 Problem Proposed . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Research Framework . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Object Tracking 9
2.1 Haar-like Features . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 AdaBoost Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 A Cascade of Classifier . . . . . . . . . . . . . . . . . . . . . . . 14
3 Transfer Learning, On-line Learning, and Adaptive Learning 18
3.1 Transfer Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.1 Transfer AdaBoost (TrAdaBoost) as Transfer Learning
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1.2 TrAdaBoost Implementation . . . . . . . . . . . . . . . . 22
3.2 On-line Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Adaptive Learning . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4 Transfer Learning, On-line Learning, and Adaptive Learning for
Object Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4 Experiment Results 32
4.1 Data Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Experiment 1: TrAdaBoost to Select the Useful Information
from Source Domain . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 Experiment 2: The Effectiveness of Transfer Learning . . . . . 38
4.3.1 The View is Changing . . . . . . . . . . . . . . . . . . . 40
4.3.2 The View and the Size are Changing . . . . . . . . . . . 40
4.3.3 Occlusion, Illumination, and the View Changing . . . . . 45
5 Conclusion and Future Work 51
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

[1] A. Blum. On-line algorithms in machine learning. In Developments from
a June 1996 seminar on Online algorithms: the state of the art, pages
306–325, London, UK, 1998. Springer-Verlag.
[2] A. Blum. Machine learning theory. Technical report, Department of
Computer Science, Carnegie Mellon University, 2008.
[3] D. Comaniciu, V. Ramesh, and P. Meer. Real-time tracking of non-rigid
objects using mean shift. In Computer Vision and Pattern Recognition,
2000. Proceedings. IEEE Conference on, volume 2, pages 142 –149 vol.2,
2000.
[4] W. Dai, Q. Yang, G.-R. Xue, and Y. Yu. Boosting for transfer learning.
In Proceedings of the 24th international conference on Machine learning,
ICML ’07, pages 193–200, New York, NY, USA, 2007. ACM.
[5] A. Ferreira. Survey on Boosting Algorithms for Supervised and Semi-
supervised Learning. In WWW, 2007.
[6] Y. Freund and R. E. Schapire. A short introduction to boosting. In In
Proceedings of the Sixteenth International Joint Conference on Artificial
Intelligence, pages 1401–1406. Morgan Kaufmann, 1999.
[7] D. Geronimo. Haar-like features and integral image representation. Tech-
nical report, Universitat Autonoma de Barcelona, 2009.
[8] H. Grabner and H. Bischof. On-line boosting and vision. In Proceedings
of the 2006 IEEE Computer Society Conference on Computer Vision and
Pattern Recognition - Volume 1, pages 260–267, Washington, DC, USA,
2006. IEEE Computer Society.
[9] A. Jepson, D. Fleet, and T. El-Maraghi. Robust online appearance models
for visual tracking. In Computer Vision and Pattern Recognition, 2001.
CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference
on, volume 1, pages I–415 – I–422 vol.1, 2001.
[10] T.-K. Kim, T. Woodley, B. Stenger, and R. Cipolla. Online multiple
classifier boosting for object tracking. In Computer Vision and Pattern
Recognition Workshops (CVPRW), 2010 IEEE Computer Society Confer-
ence on, pages 1 –6, june 2010.
[11] T. Morimoto, O. Kiriyama, Y. Harada, H. Adachi, T. Koide, and H. Mat-
tausch. Object tracking in video pictures based on image segmentation
and pattern matching. In Circuits and Systems, 2005. ISCAS 2005. IEEE
International Symposium on, pages 3215 – 3218 Vol. 4, may 2005.
[12] M. Oren, C. Papageorgiou, P. Sinha, E. Osuna, and T. Poggio. Pedestrian
detection using wavelet templates. In in Computer Vision and Pattern
Recognition, pages 193–199, 1997.
[13] S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Trans. on
Knowl. and Data Eng., 22:1345–1359, October 2010.
[14] S. Shalev-Shwartz. Online learning: Theory, algorithms, and application.
Technical report, Hebrew University, 2007.
[15] S. Stalder, H. Grabner, and L. van Gool. Beyond semi-supervised track-
ing: Tracking should be as simple as detection, but not simpler than
recognition. In Computer Vision Workshops (ICCV Workshops), 2009
IEEE 12th International Conference on, pages 1409 –1416, 27 2009-oct.
4 2009.
[16] F. Tang, S. Brennan, Q. Zhao, and H. Tao. Co-tracking using semi-
supervised support vector machines. In Computer Vision, 2007. ICCV
2007. IEEE 11th International Conference on, pages 1 –8, oct. 2007.
[17] P. Viola and M. Jones. Robust real-time object detection. In International
Journal of Computer Vision, 2001.