Motion Capture Behavior Recognition via
Neighborhood Preserving Dictionary Learning
Gao-Feng He, Shu-Juan Peng
Department of Computer Science and Technology
Huaqiao University, Xiamen, P.R. China
Email: zziahgf@qq.com, pshujuan@hqu.edu.cn
Xin Liu
∗
Department of Computer Science and Technology
Huaqiao University, Xiamen, P.R. China
∗
Corresponding Author; Email: xliu@hqu.edu.cn
Abstract—Behavior recognition from large available motion
capture data has received wide attention in the computer anima-
tion community and is growing increasingly important in recent
years. In this paper, we present an efficient motion capture behav-
ior recognition approach via neighborhood preserving dictionary
learning. First, we normalize all the motion sequences in the
database to make the motion to be comparable. Then, the neigh-
borhood preserving property is exploited using Iterative Nearest
Neighbors algorithm and subsequently added as a constraint
condition for discriminative dictionary learning, whereby the raw
motion frame can be represented as a compact set of atoms
consisting of neighborhood preserving characteristics. Finally, the
recognition result can be efficiently obtained by sparse coding
based classification scheme. Extensive experiments tested on
publicly available motion capture databases have demonstrated
the accuracy and effectiveness of the proposed approach.
Index Terms—Behavior recognition, motion capture data,
neighborhood preserving property, discriminative dictionary
learning, Iterative Nearest Neighbor
I. INTRODUCTION
Real-time motion capture systems aiming to provide a
precise representation of the complex human or other object
motions, have recently sparked a revolution in the computer an-
imation industry. With the increasing collection of available 3D
motion capture data, behavior recognition from large human
motion capture sequences has raised a considerable interest
because of its attractive applications ranging from computer
animation, virtual reality applications, 3D film production,
sports biomechanics, athletic training and so forth.
Human motion capture data recorded by a large number
of mechanical degrees of freedom is able to provide a highly
precise representation of the real body movements, and the
behavior motions with specific semantic can be popularly
utilized for a particular character simulation. In general, the
motion capture behavior recognition can be stated as the pro-
cess of automatically labeling a motion sequence with respect
to the depicted motions, which mainly comprises of two key
problems: feature representation and semantic recognition. For
the former case, the discriminative representation about the
spatial and temporal characteristics of motion sequences can
be used to increase the separability of different motions. In
the past, much effort has been done to extract representative
features and many different feature representations have been
designed to characterize the human behaviors, e.g., the se-
quences of the most informative joints [1], a dynemes and
forward differences representation [2], joint-angle rotations
of the important joints [3] and joint distance matrix repre-
sentation [4]. For the latter case, the classification scheme
incorporating the training process was specifically designed
to differentiate the diverse motions. The key issue of the
classification is to explore the most effective method for
similarity discrimination, typical algorithms include dynamic
time warping algorithm [5], histogram-based classification
method [6], decision forests [7] and sparse coding [8]. Al-
though different classification methods have been exploited to
achieve motion recognition, it is still a challenging task to
precisely recognize the human behavior due to the articulated
complexity and diversity within the whole motion capture
sequence. For instance, the similar motion clips sharing the
similar movements often appear within the different behaviors,
which often confuse the recognition process.
Recently, sparse coding has been proved to be an extremely
successful tool for finding a compact signal representation [9],
[10], [11]. Inspired by this finding, in this paper, we mainly
focus on studying the behavior recognition problem from
the human motion capture data and thus present an efficient
recognition approach via neighborhood preserving dictionary
learning. The main contributions of this work are two-fold:
1) The neighborhood preserving property within each motion
frame is well exploited, through which the frame relationships
within each pose can be well utilized to characterize the behav-
ior; 2) The dictionary learning with neighborhood preserving
constraint holds a higher discrimination power to distinguish
the articulated complexity of diverse motions.
The rest of this paper is organized as follows: Section II
will briefly survey the related work, and Section III compre-
hensively introduces the proposed approach. In Section IV, the
experiments tested on public available database and compar-
isons with existing methods are presented. Finally, we draw a
conclusion in Section V.
II. RELATED WORK
Behavior recognition is one of the most popular problems in
the pattern recognition community. In the past, a few specific
researches have been conducted on motion capture behavior
978-1-4799-8697-2/15$31.00 (c)2015 IEEE