Optical Flow Statistics for Violent Crowd Behavior Detection
Pallavi D. Chakole
a
and Vishal R. Satpute
b
Image Processing and Computer Vision Laboratory, Department of Electronics and Communication Engineering,
Visvesvaraya National Institute of Technology, Nagpur, Maharashtra, 440010, India
Keywords:
Human Activity Detection, Crowd Behavior Analysis, Anomaly Detection, Optical Flow, Similarity Index,
Correlation Coefficients.
Abstract:
We proposed an approach for identifying human violent behavior by evaluating the optical flow of a series
of sequences obtained from a video. The term Violent or Violence refers to an event that arises, causing of
unexpected displacement of a crowd. “Crowd Behaviour Analysis” is an important research topic that falls un-
der the area of image processing and computer vision, machine learning, and deep learning, which have been
investigated by researchers. Proceeding with this attitude, a simple and novel method based on the amount
of movement present in the current frame with respect to its previous frame has to be presented here. The
methodology employed is as follows: the optical flow of two consecutive frames will be calculated. Fur-
ther, correlation coefficients will be calculated by considering the magnitude of the optical flow of successive
frames. From those correlation values, we can know how much the successive frames are similar or correlated.
High correlation coefficients pointed that, there will be less movement in the crowd, a lower rate of change of
velocity, and thus normal behavior or non-violent event. On contradictory if the correlation coefficients seem
to be low, there will be more movement in the crowd, a high rate of change of velocity, and thus abnormal
behavior or violent event detected. Decision criteria have to be set for a particular threshold value that has
been selected adaptively, below which we can get violent events. Implementation has to be done on MATLAB
R2021b, using the UMN video dataset consisting of 11 videos of three different scenarios. Evaluation results
concluded that the proposed methodology can able to detect violent anomalies somehow accurately.
1 INTRODUCTION
We are constantly insecure as the population grows,
human behavioural elements diversify, and highly
crowded settings emerge all around us. People require
a security guard to counteract this insecurity. How-
ever, keeping a constant eye out for suspicious activ-
ity, especially in crowded areas, is a difficult assign-
ment for a guard. As a result of this, surveillance cam-
eras are being developed. However, there are a few
flaws. When it comes to CCTV surveillance, we save
real-time recordings in our database, and anytime
something suspicious occurred, we searched these
databases for reasons and actions. Finding anomalies
in a busy environment, however, remains difficult.As
a result, it is critical to design a real-time suspicious
activity detection system or anomaly detection sys-
tem for constantly monitoring crowded places, crowd
management, and preventing anomalies in advance.
a
https://orcid.org/0000-0001-7714-2005
b
https://orcid.org/0000-0001-9944-9489
In response to these disadvantages, researchers
and practitioners in the fields of image processing and
computer vision, machine learning, and deep learn-
ing devote themselves to detecting such anomalies
in human behaviour. Detecting unusual humanity in
public locations is a significant challenge that every-
one must investigate. Crowds are common in pub-
lic places such as public places, financial institutions,
and roadways in the modern era. Again, large crowds
may attend public events such as gatherings, concerts,
sports, protests, and demonstrations. There is always
a systemic risk, insecurity, and management required.
However, gathering management and security appear
to be ineffective.We are always looking for isolated
incidents in real-time, but all we have are recorded
videos. When something unexpected occurs in public
places, we go into these stored database systems to
figure out what happened. How did it happen? Who
should be held accountable? However, the damage
had already been done at this point. We always glance
for an automotive crowd behavioral assessment sys-
tem that takes all of these factors into account.
Chakole, P. D. and Satpute, V. R.
Optical Flow Statistics for Violent Crowd Behavior Detection.
DOI: 10.5220/0013628300004664
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 557-564
ISBN: 978-989-758-763-4
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
557

Paper is organized in a sequence as, literature re-
view, proposed methodology, result analysis, conclu-
sion and at last references are listed.
2 LITERATURE REVIEW
Researchers must be able to recognize human behav-
ior from video sequences. Behavior recognition is
concerned not only with human behaviors but also
with their mental capacities or psychological state, as
well as their facial expressions and gestures. Due to
several real-time issues such as occlusion, crowded
surroundings, illumination effect, scaled variations,
changes in views, and so on, recognizing actions from
movies or steady frames is once again a difficult task
(Parate et al., 2018). It’s a multimodal investigation
of human activity recognition. (Vrigkas et al., 2015)
proposes unimodal and multimodal modalities-based
work. The novel expresses how the influence of hu-
man behavior is dependent not just on individuals’ be-
havior but also on group behavior, their interactions,
associations, body language, and interaction with ob-
jects. By differentiating the terms ‘Activity’ as a se-
quence of actions related to body part movements,
and ‘Behavior’ as a movement of body parts along
with facial expressions, moods, and gestures. The
hierarchical structure of human activities has been
structured with respect to different layers in (Zhang
et al., 2017). Layers are categorized from bottom to
top as ‘action primitives layer’, ‘action/activity layer’,
and ‘complex interactions layer’. Adopting a different
approach to detect activities of human Kinect sensors
has been used (Nale et al., 2021) which can able to
extract features as per the user’s need. Skelton-based
unusual activities can be detected by (Franco et al.,
2020). Another approach to extract features to detect
suspicious human activities has been implemented in
(Abrishami Moghaddam and Zare, 2019) by defining
the term ‘spatiotemporal wavelet correlogram’ work
based on correlation of wavelet coefficients with re-
spect to space and time. This method is overcome
for the background subtraction method to detect fore-
ground objects. The term, ’Groups’ and ‘Crowd’ are
defined based on the number of individuals involved,
interactions within an individual, their velocity, and
direction of movements. According to (Murino et al.,
2017) ‘Groups’ are those where two or more people
are involved with some interactions, share the same
spatial and temporal adomain, as well as same veloc-
ity and direction of movements. On the other hand,
‘Crowds’ are those where more than two individuals
are involved, with no interaction, with different spa-
tial and temporal domains, with different velocities
and directions of movement. Based on the defini-
tion of groups and crowds the evaluation of activity
detection has been done by means of adopting two
different approaches as ‘Microscopic’ and ‘Macro-
scopic’ approaches (Bour et al., 2019). The micro-
scopic method is concerned with groups in which
each individual behavior is treated separately.In jux-
taposition, the macroscopic approach is concerned
with crowds, where the entire crowd is regarded as
a single entity. A summary of an anomalous group
and crowd behavior analysis has been summarised
in (Afiq et al., 2019) (Sawarbandhe et al., 2019).
In his review authors differentiates various detection
techniques that have been implemented in the last 5
year. The methods has been categorised and sub-
categorized as ‘Gaussian of Mixture Model (GMM)’
(Chavan et al., 2018)(Naveen et al., 2014), ‘Hidden
Markov Model (HMM)’ (Satpute et al., 2014)(Gan-
gal et al., 2014a)(Gangal et al., 2014b), ‘Optical
Flow (OF)’ (Nayan et al., 2019)(Chen and Lai, 2019)
and ‘Spatio-temporal techniques (STT)’. Optical flow
with the multiresolution concept has been proposed
in (Meinhardt-Llopis and S
´
anchez, 2012). Addition-
ally, the work-based of people aggregation based on
groups and crowd has been explained in (Mohammadi
et al., 2016b). In this work, the authors categorized
behavior analysis strategy as ‘model-based strategy’
and ‘motion-based strategy’. Under the model-based
approach, some algorithm has to de learned by the
model to do a specific task as per users’ requirement,
such as ‘Social Force Model (SFM)’ (Mehran et al.,
2009), ‘Behavior Heuristic Model (BHM)’ (Moham-
madi et al., 2016a). Another approach motion-based
behavior analysis has been done by means of op-
tical flow vectors (Horn and Schunck, 1981) (Lu-
cas et al., 1981), which can able to gives informa-
tion based on the amount of motion has to de done
frame by frame in a video. for example, violence
flow method (Hassner et al., 2012), the substantial
derivative method based on fluid mechanics (Moham-
madi et al., 2015), optical flow analysis based on div-
curl(Chen and Lai, 2019), entropy (Cheggoju et al.,
2021), correlation coefficients (Nayan et al., 2019),
histogram of magnitude and momentum (Bansod and
Nandedkar, 2020). Detection and classification has
been done by different techniques by different au-
thors like, SVM (Patil and Biswas, 2017)(Thombare
et al., 2021), Bag of Visual Words (BoW)(Sharma
and Dhama, 2020), tracking techniques (Jirafe et al.,
2021), (Ghutke et al., 2016). An ‘Intelligent video
surveillance’ (Sreenu and Durai, 2019) deep learn-
ing method-based review for crowd behavior analysis,
object detection(Gajbhiye et al., 2017), violent de-
tection has been summarized from different journals
INCOFT 2025 - International Conference on Futuristic Technology
558

and years, along with challenges (Gupta et al., 2016),
motivations (Pawade et al., 2021), applications, draw-
backs, and results.
3 PROPOSED METHODOLOGY
The methodology has been proposed shown in the fig-
ure.1, starting with taking a video as an input. From
video, frames are extracted for further optical flow
analysis. Optical flow is the method that can able
to give information about movements present within
successive frames in the form of motion vectors (mo-
tion in x and y direction), magnitude of motion vec-
tors, and orientation of motion. The further process
has been done by considering the magnitude of op-
tical flow only. Correlation can be estimated be-
tween two successive frames. Correlation can give
how much the two frames are similar to each other.
For smaller motion, we get large correlation coeffi-
cients and for larger motion, we get small correlation
coefficients. Further differences of correlation coef-
ficients are calculated to know about the amount of
motion. Depending on the value of differences of the
magnitude of correlation coefficients of the succes-
sive frame we get the frame number from which mo-
tion gets start increases. Accordingly, from that frame
number, we can select the correlation coefficient value
as a threshold value to set decision criteria. By com-
paring with this selected threshold we can able to de-
tect violent and non-violent events successfully. Step
by step process is explained below in detail.
3.1 Background Subtraction
As shown in figure 1, initially CCTV video is taken as
an input. For the CCTV system before getting started
initial frame is captured as the background reference
frame. From that reference frame, each extracted cur-
rent frame gets subtracted to get foreground objects
as shown in figure2.
O(x, y) = C(x, y)− B(x, y) (1)
Where O(x, y) is the foreground object frame, that we
got by differences of C(x, y) current frame, and B(x, y)
background frame for corresponding pixels coordi-
nates x and y.
3.2 Optical Flow
As an estimation of motion-based approach method
based on optical flow, Horn & Schunk (Horn and
Schunck, 1981), Lucas & Kanade (Lucas et al., 1981)
methods are well known. Optical flow is able to de-
tect movements within consecutive frames. Optical
flow is a global estimation, which means calculations
are done pixel by pixel. It is based on some assump-
tions like ‘Brightness constancy, Spatial coherence,
and Temporal persistence’. By means of brightness
constancy, the brightness of the small area remains
unchanged even if, the area gets displaced by a small
amount. Due to the property of special coherence,
the velocity of nearby points in the scene is usually
the same since they belong to the same surface. Over
time, the picture motion of a surface patch varies due
to temporal persistency.
For a video sequence I(x,y,t), by assuming bright-
ness constancy, pixel intensities remains same over
time ie.
I(x, y, t) = I(x + dx, y + dy, t + dt) (2)
By applying Taylor series expansion, some sort of
substitutions, and Crammer’s rule on eq (2) can be
written in the form of,
I
x
dx
dt
+ I
y
dy
dt
+ It = 0 (3)
Further substitutions for
dx
dt
= u and
dy
dt
= v as motion
vectors for x and y directions respectively in eq (3),
I
x
u + I
y
v + I
t
= 0 (4)
This is nothing but an optical flow constraint equation
for motion vectors u and v respectively for x and y
directions.
Further u and v are calculated by iterative meth-
ods applied on minimized cost function of brightness
constancy term along with the smoothness factor.
u
n+1
= u
-n
− I
x
(
I
x
u
-n
+ I
y
v
-n
+ I
t
α
2
+ I
x
2
+ I
y
2
) (5)
v
n+1
= v
-n
− I
y
(
I
x
u
-n
+ I
y
v
-n
+ I
t
α
2
+ I
x
2
+ I
y
2
) (6)
where n is no of iterations, u
-n
is local average ve-
locity of u for an n
th
iteration, which can be used for
subsequent iteration i.e. n + 1 iteration. Simillarly v
-n
is local average velocity of v for an n
th
iteration. α is
the scale factor to be used for smoothness constraint.
The number of iterations gets stopped if con-
vergence happened or stopping criteria gets ful-
filled, whichever gets earlier (Meinhardt-Llopis and
S
´
anchez, 2012). Further magnitude and orientation
can be calculated by using u and v getting from eq.(5)
and eq.(6) as
Mag =
p
u
2
+ v
2
(7)
θ = tan
-1
(
v
u
) (8)
Optical Flow Statistics for Violent Crowd Behavior Detection
559

Figure 1: Flow for proposed methodology
(a) (b)
(c)
Figure 2: Background Subtraction, a)Background Frame, b)Current Frame, c)Foreground object frame.
As of now by applying optical flow on image
sequences we can know about motion present in a
present image with respect to the previous image.
u, v, Mag, θ can able to give information about mo-
tion in the x-direction, motion in the y-direction, total
motion, and direction where motion happened respec-
tively. This information can be used further for pre-
dictions about events.
3.3 Correlation Coefficients
By considering the optical flow magnitude of a series
of sequences we can find out the correlation coeffi-
cients between the present image and the next consec-
utive image. Correlation coefficients can tell about,
how much similar kind of motion that the images
have? Higher the value of the coefficient, similar or
slow in motion. Lower the value of coefficients,fast in
motion. Correlation coefficients of an image I
n
with
respect to I
n+1
can be calculated by eq.(9).
R
n,n+1
=
∑ ∑
(I
n
(x,y)−µ
n
)
2
(I
n+1
(x,y)−µ
n+1
)
2
q
(
∑ ∑
(I
n
(x,y)−µ
n
)
2
)(
∑ ∑
(I
n+1
(x,y)−µ
n+1
)
2
)
(9)
Here in eq.(9) R used to represent correlation co-
efficients between image I
n
and I
n+1
. µ
n
is the mean
of image I
n
and µ
n+1
is the mean of image I
n+1
.
As we know correlation coefficients give an idea
about how much motion is present inside each im-
age. Here in this paper events may be considered as
sudden and quick movements within a crowd. That
can be observed by calculating differences of consec-
utive correlation coefficients. Smaller the differences,
slower movements within a crowd, larger the differ-
ences, large movements within a crowd. By evaluat-
ing these differences we can know the region where
anomaly exists.
3.4 Decision Criteria
Differences in correlation coefficients values give an
idea about the region where the event occurrences
can happen. A threshold value can be selected for
making a decision about an event, that the value of
correlation coefficients of that frame from where
differences of correlation coefficients start increas-
ing. For a selected threshold value we can compare
correlation coefficients of all images. If the value of
the correlation of an image sequence is less than the
selected threshold then a non-violent event can be
detected. But if the value of correlation coefficients
of a particular sequence is larger than a selected
threshold then we can say that a violent event can be
detected.
INCOFT 2025 - International Conference on Futuristic Technology
560

Figure 3: UMN dataset frames, row one is of normal event,
row two is of Abnormal event frame.
R ≤ T hreshold ; Violent event detected
Otherwise ; Non-Violent event detected
4 RESULT ANALYSIS
The implementation has been applied on UMN
dataset (Bansod and Nandedkar, 2020; Chen and Lai,
2019; Nayan et al., 2019). The simulation was done
on a CPU with 32GB of RAM, a 3.40GHz Intel(R)
Core(TM) processor.
4.1 Dataset
The dataset was generated by the University of Min-
nesota, especially for unusual activity detection for
both indoors and outdoor activities. The video dataset
consists single video combination of eleven videos of
three different weather scenarios as perfect daylight,
perfect illumination effect, and poor illumination ef-
fects. The resolution of video is 320 by 240 pixels,
with frame rate of 30fps. Some sample frames are
shown in the figure 3. For evaluation purposes, a sin-
gle video gets separated into a total of 11 different
videos of a 30fps frame rate. As shown in the table.
1 no of frames each video consists and frame number
at which event gets detected are tabulated.
Figure 4: Plot of correlation coefficients of optical flow
magnitude versus frame number
4.2 Analysis on Dataset
Test video no.-2: A very second video of the UMN
dataset is of outdoor activity video with slightly low
illumination. Video is having a total of 828 frames of
resolution 320 by 240 pixels. The proposed method-
ology was implemented on the video to get results for
violent detection. Simulation graphs are shown in fig-
ure.4 and 5 of correlation vs frame number and event
detected graph along with threshold line respectively.
Valley is the region where a sudden fall of correlation
coefficients happened due to quick movements within
a crowd. A threshold value can be selected adaptively
by means of using a difference of correlation coeffi-
cients. Depending upon the selected threshold value
event can be detected as a violent or non-violent event
successfully in figure 6.
Figure 5: Threshold plot for an video with threshold se-
lected at th=0.9480
Figure 6: Event detection plot with violent and non-violent
event regions
Test video no.-4: The fourth video of the UMN
dataset of indoor activities with poor illumination of
duration 22 sec, with a frame rate of 30fps, each frame
is of the size of 320 x 240. The simulation plots for
a video are shown in figure.7,8, and 9 are of corre-
lation plot, threshold plot, and event detection plot
respectively. Further, the evaluation of all videos of
the UMN dataset has been tried over algorithm. The
table 1 gives the statistics about number of frames, se-
Optical Flow Statistics for Violent Crowd Behavior Detection
561

Table 1: Evaluation table for UMN dataset
Video No. No. of frames
in Each Video
Simulation
Time (sec)
Threshold
Value (th)
Event Occurring
at frame No.
Ground truth
Video 1 625 24.69 0.9082 495 485
Video 2 828 35.86 0.9480 675 679
Video 3 549 19.43 0.9948 323 320
Video 4 685 27.47 0.9966 573 674
Video 5 768 31.97 0.9940 502 496
Video 6 579 20.42 0.9976 453 460
Video 7 895 41.02 0.9948 743 741
Video 8 667 25.19 0.9968 463 469
Video 9 658 25.00 0.8369 569 552
Video 10 677 25.93 0.8090 599 578
Video 11 808 33.39 0.8954 739 725
lected threshold, and frame no at which event can be
detected. It can be observed that event detected frame
is very close to ground truth frame number.
Figure 7: Plot of correlation coefficients of optical flow
magnitude versus frame number
Figure 8: Threshold plot for an video with threshold se-
lected at th=0.9966
Figure 9: Event detection plot with violent and non-violent
event regions
5 CONCLUSION
This paper is all about the optical flow method used
to detect violent events within a crowded scene video
dataset. Violent is referred to as an instant at which
people start dispersing suddenly or quick dispersal of
people within the crowd. Correlation coefficients are
used to analyze the relationship between successive
frames so that to know about the number of move-
ments. If successive frames have high correlation
means frames have low motion, whereas if correla-
tion coefficients are low then motion within the frame
is large. Here we need to target the frames where we
get large motions as large motions pointed towards
the quick and sudden action hence violent event.
REFERENCES
Abrishami Moghaddam, H. and Zare, A. (2019). Spatiotem-
poral wavelet correlogram for human action recogni-
INCOFT 2025 - International Conference on Futuristic Technology
562

tion. International Journal of Multimedia Information
Retrieval, 8(3):167–180.
Afiq, A., Zakariya, M., Saad, M., Nurfarzana, A., Khir, M.
H. M., Fadzil, A., Jale, A., Gunawan, W., Izuddin,
Z., and Faizari, M. (2019). A review on classifying
abnormal behavior in crowd scene. Journal of Visual
Communication and Image Representation, 58:285–
303.
Bansod, S. D. and Nandedkar, A. V. (2020). Crowd
anomaly detection and localization using histogram
of magnitude and momentum. The Visual Computer,
36(3):609–620.
Bour, P., Cribelier, E., and Argyriou, V. (2019). Crowd be-
havior analysis from fixed and moving cameras. In
Multimodal Behavior Analysis in the Wild, pages 289–
322. Elsevier.
Chavan, R., Gengaje, S., and Gaikwad, S. (2018). Multi-
object detection using modified gmm-based back-
ground subtraction technique. In International Con-
ference on ISMAC in Computational Vision and Bio-
Engineering, pages 945–954. Springer.
Cheggoju, N., Nawandar, N. K., and Satpute, V. R. (2021).
Entropy: A new parameter for image deciphering. In
Proceedings of International Conference on Recent
Trends in Machine Learning, IoT, Smart Cities and
Applications, pages 681–688. Springer.
Chen, X.-H. and Lai, J.-H. (2019). Detecting abnormal
crowd behaviors based on the div-curl characteristics
of flow fields. Pattern Recognition, 88:342–355.
Franco, A., Magnani, A., and Maio, D. (2020). A multi-
modal approach for human activity recognition based
on skeleton and rgb data. Pattern Recognition Letters,
131:293–299.
Gajbhiye, P., Naveen, C., and Satpute, V. R. (2017).
Virtue: Video surveillance for rail-road traffic safety at
unmanned level crossings;(incorporating indian sce-
nario). In 2017 IEEE Region 10 Symposium (TEN-
SYMP), pages 1–4. IEEE.
Gangal, P., Satpute, V., Kulat, K., and Keskar, A. (2014a). A
novel approach based on 2d-dwt and variance method
for human detection and tracking in video surveil-
lance applications. In 2014 International Conference
on Contemporary Computing and Informatics (IC3I),
pages 463–468. IEEE.
Gangal, P., Satpute, V., Kulat, K., and Keskar, A. (2014b).
Pnew approch for object detection and tracking using
2d-dwt. In 2014 International Conference on Elec-
tronics and Communication Systems (ICECS), pages
1–5. IEEE.
Ghutke, R. C., Naveen, C., and Satpute, V. R. (2016).
A novel approach for video frame interpolation us-
ing cubic motion compensation technique. Inter-
national Journal of Applied Engineering Research,
11(10):7139–7146.
Gupta, A., Stapute, V., Kulat, K., and Bokde, N. (2016).
Real-time abandoned object detection using video
surveillance. In Proceedings of the International Con-
ference on Recent Cognizance in Wireless Communi-
cation & Image Processing, pages 837–843. Springer.
Hassner, T., Itcher, Y., and Kliper-Gross, O. (2012). Violent
flows: Real-time detection of violent crowd behav-
ior. In 2012 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition Workshops,
pages 1–6. IEEE.
Horn, B. K. and Schunck, B. G. (1981). Determining optical
flow. Artificial intelligence, 17(1-3):185–203.
Jirafe, A., Jibhe, M., and Satpute, V. (2021). Camera hand-
off for multi-camera surveillance. In Applications
of Advanced Computing in Systems, pages 267–274.
Springer.
Lucas, B. D., Kanade, T., et al. (1981). An iterative image
registration technique with an application to stereo vi-
sion. Vancouver.
Mehran, R., Oyama, A., and Shah, M. (2009). Abnormal
crowd behavior detection using social force model. In
2009 IEEE conference on computer vision and pattern
recognition, pages 935–942. IEEE.
Meinhardt-Llopis, E. and S
´
anchez, J. (2012). Horn-schunck
optical flow with a multi-scale strategy. Image Pro-
cessing on line.
Mohammadi, S., Kiani, H., Perina, A., and Murino, V.
(2015). Violence detection in crowded scenes us-
ing substantial derivative. In 2015 12th IEEE Inter-
national Conference on Advanced Video and Signal
Based Surveillance (AVSS), pages 1–6. IEEE.
Mohammadi, S., Perina, A., Kiani, H., and Murino, V.
(2016a). Angry crowds: Detecting violent events in
videos. In European Conference on Computer Vision,
pages 3–18. Springer.
Mohammadi, S., Setti, F., Perina, A., Cristani, M., and
Murino, V. (2016b). Groups and crowds: Behaviour
analysis of people aggregations. In International Joint
Conference on Computer Vision, Imaging and Com-
puter Graphics, pages 3–32. Springer.
Murino, V., Cristani, M., Shah, S., and Savarese, S. (2017).
Group and crowd behavior for computer vision. Aca-
demic Press.
Nale, R., Sawarbandhe, M., Chegogoju, N., and Satpute,
V. (2021). Suspicious human activity detection us-
ing pose estimation and lstm. In 2021 Interna-
tional Symposium of Asian Control Association on In-
telligent Robotics and Industrial Automation (IRIA),
pages 197–202. IEEE.
Naveen, C., Satpute, V. R., Kulat, K. D., and Keskar, A. G.
(2014). Video encoding techniques based on 3d-dwt.
In 2014 IEEE Students’ Conference on Electrical,
Electronics and Computer Science, pages 1–6. IEEE.
Nayan, N., Sahu, S. S., and Kumar, S. (2019). Detecting
anomalous crowd behavior using correlation analysis
of optical flow. Signal, Image and Video Processing,
13(6):1233–1241.
Parate, M. R., Satpute, V. R., and Bhurchandi, K. M. (2018).
Global-patch-hybrid template-based arbitrary object
tracking with integral channel features. Applied In-
telligence, 48(2):300–314.
Patil, N. and Biswas, P. K. (2017). Detection of
global abnormal events in crowded scenes. In 2017
Twenty-third National Conference on Communica-
tions (NCC), pages 1–6. IEEE.
Optical Flow Statistics for Violent Crowd Behavior Detection
563

Pawade, A., Anjaria, R., and Satpute, V. (2021). Suspi-
cious activity detection for security cameras. In Ap-
plications of Advanced Computing in Systems, pages
211–217. Springer.
Satpute, V., Kulat, K., and Keskar, A. (2014). A novel ap-
proach based on 2d—dwt and variance method for hu-
man detection and tracking in video surveillance ap-
plications (an alternative approach for object detec-
tion). In 2014 9th International Conference on In-
dustrial and Information Systems (ICIIS), pages 1–6.
IEEE.
Sawarbandhe, M. D., Satpute, V. R., Cheggoju, N., and
Sinha, S. (2019). Combining learning and hand-
crafted based features for face recognition in still im-
ages. In TENCON 2019-2019 IEEE Region 10 Con-
ference (TENCON), pages 2502–2506. IEEE.
Sharma, S. and Dhama, V. (2020). Abnormal human be-
havior detection in video using suspicious object de-
tection. In ICDSMLA 2019, pages 379–388. Springer.
Sreenu, G. and Durai, M. S. (2019). Intelligent video
surveillance: a review through deep learning tech-
niques for crowd analysis. Journal of Big Data,
6(1):1–27.
Thombare, P., Gond, V., and Satpute, V. (2021). Artificial
intelligence for low level suspicious activity detection.
In Applications of Advanced Computing in Systems,
pages 219–226. Springer.
Vrigkas, M., Nikou, C., and Kakadiaris, I. A. (2015). A
review of human activity recognition methods. Fron-
tiers in Robotics and AI, 2:28.
Zhang, S., Wei, Z., Nie, J., Huang, L., Wang, S., and Li, Z.
(2017). A review on human activity recognition using
vision-based method. Journal of healthcare engineer-
ing, 2017.
INCOFT 2025 - International Conference on Futuristic Technology
564