PHSW-based automated human behavior segmentation X. Weiwei
et al.
Figure 1. Behavior segmentation.
features. Moreover, a selective combination of posture fea-
tures is proposed for the first time to derive the posture
histogram. To reduce noise sensitivity, we try to analyze
behavioral features in subsequence level by applying slid-
ing window. And in order to obtain conspicuous and stable
behavioral characteristics of the extracted subsequences, a
novel way is put forward to tune sliding window by ana-
lyzing steady states of human behaviors. Benefiting from
the great clustering property of posture histograms in slid-
ing window, behavior segmentation problem is tactfully
simplified to the detection of outlier subsequence whose
central frame is taken as the segmenting point. And the
local outlier factor (LOF) algorithm [26] is first adopted
to pick up the outlier subsequences. CMU MoCap dataset
is used as our dataset for the analysis and experiments.
As shown in the experiment results, our proposed method
achieves great performance in precision and recall, which
outperforms other state-of-the-art methods. The main con-
tributions of this paper can be concluded as follows:
(1) A set of novel features including bone–bone
angles (BBAs), joint–plane distances (JPDs), and
bone–plane angles (BPAs) is proposed and defined
to represent the human body postures. Based on
these novel features, a compact representation
of behavioral features, posture histogram,ispre-
sented for behavior segmentation. It can effectively
decrease the computation cost of segmentation
without loss of accuracy.
(2) As subsequence contains more conspicuous and sta-
ble behavioral features, we construct sliding window
and propose a novel way to tune it by analyzing
steady states of human behaviors. Thus, the behav-
ior features are analyzed in subsequence level. Com-
pared with the frame-based method, our method is
robust to noise.
(3) Regarding the behavior segmentation problem, we
tactfully simplify it to the detection of outlier sub-
sequence, and the LOF algorithm is first adopted to
solve this problem and achieves good performance.
The remainder of this paper is organized as follows.
After reviewing related works in Section 2, we construct
posture histogram from a selective feature set in Section 3.
In Section 4, we extract subsequences by a well-tuned slid-
ing window, and then the LOF algorithm is implemented to
pick up the outlier subsequences. In Section 5, experiments
are conducted on 14 multi-behavior motion sequences
from the CMU MoCap database. Finally, Section 6 con-
cludes this paper and discusses the future work.
2. RELATED WORKS
Over the past decades, several techniques have been pro-
posed for behavior segmentation, which can be generally
categorized into two types: learning and non-learning. As
for learning-based methods, they can be further divided
into two groups: supervised learning and unsupervised
learning.
Supervised learning methods generally formulize behav-
ior segmentation as a classification problem, where clas-
sifiers are trained from a carefully selected training set.
Some behavior classifiers were trained based on existing
models, such as support vector machine [ 12] and hid-
den Markov model [ 13]. In addition, some custom models
were also proposed to solve the problem [14–17]. Müller
et al. defined motion template to characterize behaviors
[14]. Barnachon et al. proposed behavior histogram as the
behavior classifier [15]. However, segmenting results heav-
ily relied on the training set, and it would fail to pick
up the segments whose corresponding behaviors were not
contained in the training set.
In order to overcome this limitation, it is natural
for researchers to use an unsupervised learning strat-
egy. In unsupervised learning methods, behavior segments
are located by clustering motion frames. Barbi
ˇ
c et al.
employed Gaussian mixture model (GMM) to model the
motion sequence [18]. The segmenting point was located
as the frame on both sides in which two sets of frames
belong to different Gaussian distributions. Yet manually
setting the number of clusters is not practical when motion
502
Comp. Anim. Virtual Worlds
2016; 27:501–514 © 2016 John Wiley & Sons, Ltd.
DOI: 10.1002/cav
评论0
最新资源