Proceedings of CCIS2018
LGMD-based Visual Neural Network for Detecting Crowd
Escape Behavior
Bin Hu
1
, Zhuhong Zhang
2*
, Lun Li
2
1
Department of Computer Science, College of Computer Science & Technology, Guizhou University, Guiyang 550025,
China
2
Department of Big Data Science & Engineering, College of Big Data & Information Engineering, Guizhou University,
Guiyang 550025, China
csuhubin@163.com, * Corresponding author, zhzhang@gzu.edu.cn, lilun@cumt.edu.cn
Abstract: Effective detection of crowd escape behavior
in public places is a challenge in computer vision.
Although some studies on such detection have been
done in recent years, no complete solution satisfies the
current application demand. This work develops an
improved neural network to detect crowd escape
behavior presented in video monitoring systems. In the
neural network, luminance change caused by crowd
activity is first collected; second, the gathered excitation
from each image pixel is mixed with the delayed
excitation from the related neighboring counterpart in a
specific proportion, relying upon the inspiration of
visual information integration in the mammalian’s retina;
finally, an unique adaptive threshold scheme is designed
to regulate the discharge excitation of the neural network
in order to perceive the process of the crowd escape
behavior. Experiments have validated that the neural
network is effective for detecting the escape behavior of
crowd, based on a public video data set of crowd escape
events.
Keywords: Event detection; Crowd escape behavior;
LGMD model; Motion perception; Video surveillance
1 Introduction
Crowd escape events are one of serious disasters in
public places [1], since it is possible to result in potential
safety hazard. Once unusual crowd activity can be
timely perceived by means of video surveillance, some
coming dangers can be controlled or alleviated. In order
to characterize and detect crowd behavior, video
monitoring systems with comprehensive engineering
applications can be designed as automatic crowd activity
perception ones without human intervention, by means
of computer vision approaches such as escape
trajectories [2], dynamic texture patterns [3], optical
flow [4], and spatio-temporal contexts[4]. However, one
such kind of system is reported rarely, due to technical
and real-time requirements. Essentially, the function of
animal vision system is similar to that of video
surveillance, as any eyed animal can all capture visual
information, e.g., objects’ motion, shape, feature, color
and so on. The captured visual cues can be converged to
execute decision making. Thus, some new inspirations,
borrowed from visual neurophysiology can give
researchers some lights on intelligent monitoring
systems.
Nature provides various source materials for different
kinds of invertebrate and vertebrate animal species to
perceive visual motion [5]–[8]. Particularly, locusts can
perceive spatio-temporal intensity change and avoid
collision with each other, for which the main reason is
because their wide-field visual neurons, so-called lobula
giant movement detectors (LGMD), share the function
of preferential selectivity for visual stimuli caused by
imminent collision [9], [10]. Locusts’ LGMD usually
excites when an object moves around their hazardous
regions rapidly. LGMD’s excitation strength is greatly
related to the intensity of the visual stimuli caused by the
moving objects in the field of view. Since it, compared
with other visual information processing neurons in
higher animals, is much simpler, its internal structure
and neural mechanism have been successfully studied in
recent years [9], [11], [12]. Especially, a LGMD-based
computational model was proposed by Rind and
Bramwell [9], while having successfully derived
multiple visual neural networks for different motion
perception tasks such as collision detection [13], [14],
translational direction detection [15], [16], rotational/
spiral motion perception [17], [18] and so on. However,
it is still open whether LGMD can be employed to
construct a simplified computational model for
autonomous video surveillance. To this point, this paper
concentrates on solving the hard issue of crowd activity
identification, by developing a LGMD-based neural
network suitable for crowd escape behavior detection
related to the application of intelligent video
surveillance. This neural network includes three layers.
The first collects luminance change caused by crowd
activity; the second, similar to the inspiration of visual
information integration [19], mixes the collected
excitation from each image pixel with the delayed
excitation from the related neighboring counterpart; the
third transmits the mixed excitation to the final output
neuron. After that, a unique adaptive threshold scheme is
designed to regulate the discharge excitation of the
neural network for the perception of the process of the
crowd escape behavior in the field of view.
The main contribution of the paper includes three points.
(i) the perception mechanisms in biological vision
systems are skillfully borrowed to depict unusual crowd
978-1-5386-6005-8/18/$31.00 ©2018 IEEE