Page 131 - Proceedings of the 2018 ITU Kaleidoscope

OPTICAL FLOW BASED LEARNING APPROACH FOR ABNORMAL CROWD ACTIVITY
DETECTION WITH MOTION DESCRIPTOR MAP

Dhananjay Kumar, Govinda Raj Sampath Sarala

Department of Information Technology, Anna University, MIT Campus, Chennai

no human detection or tracking processes involved. The
ABSTRACT interaction force is measured by computing the difference
between the desired and actual velocities obtained from the
Automated abnormal crowd activity detection with faster particle advection on the optical flow field. The major
execution time has been a major research issue in recent drawback of these systems is that the social force optical
years. In this work, a novel method for detecting crowd flow features are directly used to detect patterns, which
abnormal activities is proposed which is based on may lead to poor accuracy as even slight deviation from
processing of optical flow as motion parameter for machine normal may be detected as abnormal. Social behavior using
learning. The proposed model makes use of magnitude interaction energy potential [3] could be considered in
vector which represents motion magnitude of a block in activity detection. The interaction energy potential is
eight directions divided by a 45 degree pace angle. Further, estimated from the velocity of the space–time interest
motion characteristics are processed using Motion points to explain whether they will meet in the near future.
Descriptor Map (MDP), which takes two main parameters
namely aggregate magnitude of motion flow in a block and Deep learning technique can also be used to anomalous
Euclidean distance between blocks. Here, the angle of event detection. However, there is a need for a model that
deviation between any two blocks determines which among trains a convolutional neural network (CNN) using
the eight values in the magnitude vector to be considered spatiotemporal patches from optical flow images as input
for further processing. The algorithm is tested with two [4]. Even though these methods produce state of the art
standard datasets namely UMN and UCSD Datasets. Apart results, computation complexity is very high and hence it
from these the system is also tested with a custom dataset. takes a lot of time to process the results.
On an average, an overall accuracy of 98.08% was
obtained during experimentation. The proposed framework work (Fig. 1) to detect abnormal
behavior (event detection) in the scene, confirm to
Keywords – Optical flow, Euclidean distance, K-means, intelligent video surveillance system mentioned in ITU-T
Angle of deviation recommendation F.743.1 – “Requirements for intelligent
visual surveillance”. The output of alarm is based on
1. INTRODUCTION polygonal region demarcated by the user. The basic
requirement in F.743.1 is: when the retention time of an
In the recent years, there has been considerable interest on object in polygon area exceeds a prescribed threshold, an
computer vision to identify human activities and detect alarm needs to be triggered. The fraud detection system
abnormal scenarios from video input. It has been builds a profile of normal activity and uses a behavior
established that with the increase in the number of security analysis function to send alerts on deviations.
cameras, the efficiency and accuracy of human operators
have reached the limit. The automated system not only need
to detect and locate abnormal behavior in real time, but also
notify the agent through a report. The design and
implementation process is challenging due to lack of
specific knowledge of the scene and target activities.
Abnormal activity detection may differ from suspicious
activity detection. An activity if not been previously trained
in the system, will be recognized as abnormal. But this
activity may or may not be suspicious depending on the
human perception.

Many approaches have been adopted for global unusual
activity detection by modeling the behavior of the crowd Figure 1 - Functional components of the abnormal activity
itself. In some early approaches crowd behaviors are detection system
described by means of the Social Force model [1-2], with

978-92-61-26921-0/CFP1868P-ART @ 2018 ITU – 115 – Kaleidoscope