Instrumentation & Measurement Magazine 23-5 - 58

first phase consists of constructing the reference CNN model
that mimics the normal situations based on the data without
abnormal events (attacks). In addition, we computed the detection threshold of the kNN-DEWMA approach based on the
CNN features. In the second phase, we apply the constructed
CNN model to generated features based on testing, and we
apply the kNN-DEWMA approach to the new features using
the detection threshold previously computed to detect abnormal events.
Specifically, the CNN features are examined using the proposed kNN-DEWMA approach for the purpose of atypical
events detection. Importantly, this approach takes advantage
of the great ability of the kNN algorithm to quanify the similarity between normal and abnormal CNN's features to better
uncover atypical events. This is mainly due to its capacity to
handle nonlinear features without making hypotheses on the
data distribution.
The major reason for double exponentially smoothing
kNN measurements (kNN-DEWMA) is to include all of the
information from past and actual samples in the decision
rule, which make it sensitive to small anomalies. Moreover,
we applied the DEWMA scheme to kNN measurements
to incorporate all of the information from past and current
measurements in the decision process, which offers more
sensitivity to small changes. Furthermore, to obtain a flexible and assumption-free approach, the non-parametric
kernel density estimation is used to compute the detection

threshold. Lastly, for the spatial localization of abnormal
events in each frame of the video, the Gaussian Distribution of Mahalanobis Distance (GDMD) is applied only on
the frames flagged by the kNN-DEWMA detector. Results
on benchmark data indicate that the developed method has
a higher detection efficiency than the traditional competitors
by a large margin.

Proposed CNN-KNN for Abnormal
Attacks Detection
This section presents a vision-based method for attack detection and localization in crowded areas. The implementation
of this method is performed in two steps: (1) extracting robust
features from normal video data, based on the first two convolution layers of a pre-trained CNN; and (2) applying the
proposed KNN-DEWMA scheme to check the CNN features
to detect anomalies in the video (Fig. 1).

Features Extraction
Various model Structures of Deep Neural networks were
shown to be effective for many tasks because of their structures utilize hidden features. One of the effective tools for
features extraction in deep neural networks is CNN, and in
particular, Alexnet [12] which has been trained and learned
on a huge number of images to extract robust features from
these images. In our proposed scheme, the first step is to use
many frames: Ft, Ft-1, Ft-2, D = [Ft, Ft-1, Ft -2] as input to the

Fig. 1. Conceptual representation of the proposed monitoring method.
58

IEEE Instrumentation & Measurement Magazine

August 2020

Instrumentation & Measurement Magazine 23-5

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 23-5