Page 132 - Proceedings of the 2018 ITU Kaleidoscope

2018 ITU Kaleidoscope Academic Conference

As per the ITU-T recommendation X.1157 – “Technical to solve these problems by providing automatic detection of
capabilities of fraud detection and response for services suspicious behavior that uses contextual information. The
with high assurance level requirements”, behavior profiling idea of using context space model for abnormal activity
employs a learning phase that builds profiles of normal detection used by Xiang et al. [15] models both abnormal
activity for discrete event sources collected by monitoring and normal behaviors. The main difficulty with this system
capabilities. The proposed system is implemented in a is that all normal behaviors can’t be accurately modeled.
desktop environment and tested with standard data set and The major drawback of these works is that the system
also in a locally defined scenario, which can be used further always needed a lot of contextual training to build the
in standardization of future intelligent surveillance systems. model.

2. RELATED WORKS The Structural Context Descriptor (SCD) [16] can be used
for describing the crowd individual. By online
2.1. Statistical and Probabilistic Methods spatiotemporally analyzing the SCD variation of the crowd,
the abnormality is localized. Tracking individuals in the
In abnormal activity modeling and prediction, many crowd to detect abnormal activities performs good to
techniques are based on Gaussian Mixture Model (GMM) identify individual abnormal activities like single person
and Hidden Markov model (HMM). In [5], a multilevel running in a place where running is abnormal. But for
HMM is used to predict anomalous events in specific identifying group behaviors like panicking, convergence or
regions of the crowd. Markov models allow the analysis of divergence movements, their algorithm lags behind.
the scene [6-7]. The expectation maximization algorithm [2]
has been also employed as a predictor for anomaly. A two 2.3. Generalized Methods
levels of feature analysis and Gaussian regression process
can employed to improve performance [8]. However, a Generalized abnormal detection approach has performed
robust approach can use a hierarchical mixture of dynamic well in terms of both accuracy and efficiency. They are also
textures to describe the frame [9]. target towards implementation in real time. Approaches
such as HOFME [17] HOFM [18] for detecting anomalous
Many group modeling approaches [10-11] were proposed in events in crowded scenes used general concepts such as
recent years which used Gaussian process, codebook, bag orientation, magnitude and entropy which overcomes the
of features (BOF) etc. Some anomaly detection frameworks difficulty to create models due to their unpredictability and
used spatio-temporal system context [12]. They presented their dependency on the context of the scene. The main
instant behaviors of a single object using an atomic event, drawback of these modals is that, they consider the overall
which contained the location, sense of movement, and magnitude of the block to be processed. Here, if the
velocity of an object. Because of the huge fluctuations in momentum within the block is not unidirectional, then it
appearance, scale, lighting, and pose, it's hard to detect or may affect the accuracy by detecting wrong blocks.
track individuals in crowded scenes.
3. PROPOSED WORK
A probabilistic framework can be used for automatically
interpreting the visual crowd behavior using crowd event The proposed model is composed of training and detection
detection and classification in optical flow manifolds phases. In training phase, the normal patterns are clustered,
(OFMs)[13]. whereas during detection phase, if they differ significantly
from the normal patterns learned, they are considered as
2.2. Context based methods abnormal. There are four major modules in the system
namely optical flow of blocks computation, motion
Behaviors differ radically from one place to another, since a descriptor computation, motion descriptor pattern clustering,
human activity that is considered abnormal in one scenario and nearest neighbor search. Figure 2 shows the overall
may be normal in another. To make the system detect flow of the proposed model. The particle movements in the
context based abnormal behaviors has been an important video input are captured using optical flow. The extracted
goal to accomplish. A context based method [14] attempts optical flow is passed to a magnitude vector computing

Figure 2 - Overview of the proposed model for abnormal crowd activity detection

– 116 –