Trajectory Matching Algorithm Based on Clustering and GPR in Video Retrieval - Intelligent Computing for Sustainable Energy and Environment - page 373

Trajectory Matching Algorithm
Based on Clustering and GPR
in Video Retrieval
Tianlu Wang and Xianmin Zhang
Department of Automation, Shanghai Jiao Tong University,
Image Processing and Pattern Recognition Institution of Shanghai Jiao Tong
University, Shanghai 200240
[e-mail], [e-mail]
Abstract. This paper proposes an approach for trajectory matching in
video retrieval. Algorithm is consist of three parts. First, a terse trajec-
tory contains most important temporal and spatial characters are ab-
stracted. Then, abstracted trajectories are classified into several classes
by using reformed k-means cluster method according to their position
and acceleration features. At the end, Gaussian Process regression model
is built and trained using clustered trajectories and trajectories classes
which are similar with the given retrieval targets are found out. Ad-
vantages of this algorithm include the possibility of generalized clus-
tering for similar trajectories in different scales and partial trajectories
matching.
1 Introduction
Current methods used in pictures and videos retrieval are mostly based on text
tags, which are labeled manually. These manual operations waste too much time.
We need more automatic video processing methods and constructing a new video
retrieval system.
The application of moving object trajectory matching is adopted by many lit-
eratures. X. Li et al introduce a good two-step approach which makes clustering
time effective and calculation ecient, but there is no specific method to solve the
problem of limits on clustering trajectories with different scales and lengths [1].
O. Ossama et al adopt core-set which indicates the most essential partitions of
a motion trajectory as cluster centroid [2]. K. Kim et al introduce an important
and useful matching method which makes trajectory integrity unnecessary, but
leave a puzzle on the collection of training samples [3].
There are also many related works about content based retrieval system. Spe-
cific methods and future research direction for video retrieve system are analyzed
by W. Hu et al [4]. M. Broilo et al present the most recent motion trajectory rep-
resentation and matching methods [5]. Z. Zhang et al and R. Hu et al show us the
common useful distance measures, cluster algorithms and video data-sets [6] [7].