Performance Analysis of Different Motion Detection
Techniques.
Himani K.Borse
Bharati Patil
G.H. Raisoni College of Engg and
Mgmt.,Wagholi,Pume
G.H. Raisoni College of Engg and
Mgmt., Wagholi, Pume
Himaniborse1992@gmail.c
om
Bharti.patil@email.com
ABSTRACT
2.2 Frame Separation:
The first step of video analysis is background subtraction. In
this, we implemented the four methods and one of the method
in FPGA. Detection of foreground or moving object from
background is important task of any motion detection
techniques. We discussed the four methods as Background
subtraction; Frame difference, neural map and Background
subtraction with dirichlet process Gaussian mixture model.
From the parameters obtained from this methods, we shows
how Background subtraction with dirichlet process Gaussian
mixture model is better than others methods.
first step in the background
subtraction algorithm is the Frame processing, the purpose of
this step is to make the improved video frames by removing
noise and undesirable object’s in the frame in order to
increase the quantity of information gained from the frame
and the sensitivity of the algorithm.
Input
video
Frame
Separation
Preprocessing
Extraction
of object
Different
motion
detection
techniques
Postprocessing
Keywords
Background subtraction, Dirichlet processes, video
analysis, Frame difference, neural map,
1
INTRODUCTION
Human body motion analysis is an essential technology which
modem bio-mechanics along with computer vision and has
normally used in intelligent control, human computer
interaction, motion analysis and virtual reality and other
fields. In which the moving human body recognition is the
most significant part of the human body analysis, the purpose
is to identify moving human body from background image in
video sequence, and for the follow-up action such as the target
classification its effective recognition plays a very major role.
Three are various application of human activity in
various fields, the most significant of which is surveillance.
Other applications include advanced intelligent user
interfaces, character animation for games and movies avatars
for teleconferencing, biomechanical analysis of actions for
sports and medicine the human body tracking and behavior
understanding.
2
BLOCK
DESCRIPTION:
DIAGRAM
AND
Target
object
Figure1: Block diagram of the complete system
2.3 Pre-processing:
Preprocessing is a procedure of
collection of simple image processing tasks that alters the raw
input video into a format. This can be done by succeeding
steps. Preprocessing of the video is essential to improve the
detection of moving objects. For example, by spatial and
temporal smoothing, snow as moving leaves on a tree, should
be eliminated by morphological processing of the frames later
the identification of the moving objects
2.4 Motion detection by different
techniques: There are different techniques are
implemented for motion detection such as background
subtraction, frame difference method, neural map and
background subtraction with Gaussian mixture model.
2.1 Introduction to Digital Video: Digital video
means to the capturing, manipulation, and storing of moving
images that can be displayed on computer screens. It involves
that the moving images be digitally held using the computer.
The word digital signifies to a system built on discrete or
Discontinuous events, as opposite to a continuous, analog
event. Video can be in any format AVI, MPEG, MOV.
2.1.5. Post-Processing:
The output of foreground
detection has noise. Typically, it affects by various noise
factors. To overcome this problem of noise, it involves
additional pixel level processing.
For elimination of noise from foreground pixel map low
pass filter and morphological operations, erosion and dilation
are used. Aim in applying these operations is eliminating
noisy foreground pixels that do not resemble to actual
foreground regions, and to eliminate the noisy background
pixels close and inside object regions that are truly foreground
pixels.
2.1.6. Extraction of Moving Human Body:
Certain accurate edge regions will be got, after median
filtering and morphological operations, but the region belongs
to the moving human body could not be determined.
2.1.7. Target object: Accurate moving or foreground
object is detected form background.
3
MOTION
DETECTION
DIFFERENT TECHNIQUES:
BY
3.1 Background subtraction method:
A basic
method of detecting moving object is background
subtraction, where two frames are compared one is reference
or fixed frame and another is current frame. Pixels in the
current frame that differ significantly from the background
are detected to be moving objects. Further processing of the
“foreground” pixels are done for object localization and
tracking. Subsequently background subtraction is frequently
the major step in various computer vision applications; it’s
significant that the extracted foreground pixels perfectly
match to the motion objects of interest.
There are numerous problems that a background
subtraction algorithm must resolve appropriately. Consider a
video sequence from a static camera. As it is an outside
environment, a background subtraction algorithm must adjust
to different stages of illumination at dissimilar periods of the
day and operates on adverse weather condition like fog or
snow that changes the background. Varying shadow, cast by
moving objects, should be eliminated so that reliable features
can be collected from the objects in successive processing.
. The basis of the method is that of identifying the
moving objects from the difference of the current frame and a
reference frame, normally called the “background copy”, or
“background replica”. As the background image must be an
illustration of the scene with no motion objects and should be
kept frequently updated so as to adjust to the changing
luminance conditions and geometry settings.
Simple scheme of background subtraction is to subtract the
current image from a reference image that represents the
background scene.
Normally, the simple steps of the algorithm are as follows:




Background modeling considers a reference image
signifying the background.
Threshold selection defines suitable threshold
values used in the subtraction procedure to get a
desired detection rate.
Subtraction operation or pixel classification
categorizes the kind of a given pixel, i.e., the
difference is greater than threshold is considered as
the pixel is the part of foreground or it is a moving
object. Otherwise it is background object
If Difference > threshold foreground object
Else
background object.
3.2 Frame difference method: The aim of motion
detection is to identify motion of objects originate in the two
given images. Also, finding objects motion can corresponds to
objects identification. Therefore, the key objective of this
method is to distinguish pixels belonging to the same object.
This method gives the whole movement information and
identifies the moving entity from the background well. In the
frame subtraction method the existence of moving objects is
determined by computing the difference between two
successive images.
Detection of moving object from a sequence of frames
received from a static camera is broadly accomplished by
frame difference method. The frame difference method is the
common technique of motion recognition. It implements
pixel-based difference to find the moving object.
The frame difference method is the common method of
motion detection. This method takes on pixel-based difference
to detect the moving object.




Firstly, the first frame is taken through the static
camera and afterward that sequence of frames is
captured at fixed intervals.
Secondly, the absolute difference between the
consecutive frame and the difference image is
calculated and stored in the system
Thirdly, this difference image is changed into gray
image and then converted into binary image.
Lastly, for noise removal morphological filtering is
done.
Transformation of absolute difference image to Gray
Image: There are holes present in moving entity region, and
outline of moving object remains not closed.
The absolute differential image is converted to grey image to
simplify further-operations.
RGB to Gray: Y=0.299R + 0.587G + 0.114B
Difference of Two Successive Frames: Ik is the value of the
kth frame present in image sequences. Ik+1 is the value of the
(k+1)th frame present in image sequences. The absolute
difference image is defined as follows:
𝐼𝑑(𝐾,𝐾+1) = | 𝐼𝑘+1 – 𝐼𝑘 |
Object detection: After difference calculation, object is
detected by setting threshold value.
| 𝐼𝑘+1 – 𝐼𝑘 | > threshold
If difference is greater than threshold then foreground object
else background object.
3.3 Neural map method: This method based on selforganization over artificial neural networks, broadly useful in
human image processing systems and further commonly in
cognitive science. The proposed method can handle scenes
containing moving backgrounds, slow illumination deviations
and camouflage, has no bootstrapping limits, can contain into
the background model shadows cast by moving things, then
accomplishes robust detection for various kinds of videos
taken with stationary cameras.
The key difficulty of the background subtraction approach
to motion object recognition is its highly sensitivity to
dynamic scene variations due to lighting and extraneous
events. While they are commonly identified, they leave
behind “holes” where the recently showing background
imagery varies from the well-known background model
(ghosts). Whereas the background model ultimately adjusts to
these “holes,” they create false alarms for a small period of
time. Therefore, it is extremely required to construct an
method to motion detection built on a background model that
automatically adjusts to variations in a self-organizing style
and without a priori knowledge. This method accepts a
biologically motivated problem-solving method depends on
visual attention mechanisms. The purpose is to find the
objects that keep the user attention in accordance with a group
of predefined features, containing gray level, motion and
shape features. Each node computes a function of the
weighted linear combination of received inputs, where
weights look like the neural network learning. Doing so, each
node could be signified by a weight vector, achieved from
collecting the weights correspond to incoming links. In the
following, the group of weight vectors will be referred as a
model. An arriving pattern is mapped to the node whose
model is “most similar” (similar to a predefined metric) to the
pattern, and weight vectors in a neighbourhood of such node
are updated.
depicted in Fig. 4.1., the upper left pixel takes weight vectors
stowed into the 3*3 components of the higher left portion of
neuronal map
3.3.3 Subtraction and Update of the
Background Model: After initialization, temporally
subsequent samples are providing to the network. Each
entering pixel 𝑃𝑡 of the the 𝐼𝑡 sequence frame is matched to
the current pixel model to define if there occurs a weight
vector that greatest matches it. If a greatest matching weight
vector is establish, it means that pixel fits to the background
and it is used as the pixel encoding approximation, and the
greatest matching weight vector, together with its
neighbourhood, is reinforced. Otherwise, if no allowable
matching weight vector exists, we identified as belonging to a
moving object (foreground).
Euclidean distance of vectors gives the distance between
two pixels in the HSV color hexcone, 𝑃𝑖 =(ℎ𝑖 , 𝑠𝑖 , 𝑣𝑖 ) and
𝑃𝑗 =(ℎ𝑗 , 𝑠𝑗 , 𝑣𝑗 )
as
d(𝑃𝑖 , 𝑃𝑗 ) = ∥
(𝑣𝑖 𝑠𝑖 cos(ℎ𝑖 ), 𝑣𝑖 𝑠𝑖 sin(ℎ𝑖 )𝑣𝑖 ) −
(𝑣𝑗 𝑠𝑗 cos(ℎ𝑗 ), 𝑣𝑗 𝑠𝑗 sin(ℎ𝑗 ), 𝑣𝑗 ) ∥22
Hence, the network acts as a reasonable neural network that
implements a winner take- all function with an related
mechanism, that alters the local synaptic plasticity of the
neurons, letting learning to be limited spatially to the local
neighbourhood of the maximum active neurons.
Steps for neural map method:
1.
2.
3.
4.
5.
Take background image and apply mapping 3*3
Assign constant for thresholding depending on
environmental condition and camera
Take current image and apply 3*3 mapping
Apply distance calculation means find distance
between background images and current
image(Euclidian distance formula)
Apply thresholding and make binary image
3.3.1 Applying mapping: For each color pixel, we
assumed a neuronal map comprising of weight vectors. Every
arriving sample is mapped to the weight vector that is nearest
allowing to an appropriate distance measure, and the weight
vectors in its neighbourhood are updated..
The representation of HSV values as vectors in the HSV
color hexcone used in such distance measure avoids problems
connected with the periodicity of hue and with the instability
of hue for small values of saturation
Weight vector 𝐶𝑚 , for some m, provides the greatest
match for the entering pixel 𝑃𝑡 if its distance from is smallest
in the model C of 𝑃𝑡 and is no larger than a fixed threshold.
𝑑(𝐶𝑚 , 𝑃𝑡 ) =
lim 𝑑(𝐶𝑖 , 𝑃𝑡 ) ≤ 𝜀
𝑖=1,…,𝑛2
The threshold permits distinguishing between foreground
and background pixels, and is selected as
Fig 3..3.1(Left) simple image and (Right) Neural map
structure
3.3.2 Initial Background Model: First image of our
sequence is really a good initial estimation of the background,
and consequently, for every pixel, the corresponding weight
vectors are initializing with the pixel value. In order to signify
each weight vector, we select the HSV color space, depend on
the hue, saturation and value properties of every color. As
𝜀= {
𝜀1 ,
𝜀2 ,
𝑖𝑓 0 ≤ 𝑡 ≤ 𝐾
𝑖𝑓
𝑡≥𝐾
With 𝜀1 and 𝜀2 minor constants. Specifically, 𝜀1 must be
greater than 𝜀2 , as greater values for 𝜀1 allow, within the first
K sequence frames, to get a (possibly rough) background
model containing numerous observed pixel intensity
deviations whereas lower values for permit to 𝜀2 achieve a
more correct background model in the online phase.
3.3.4 Updating in the Neighbourhood: In Fig. 4.1,
Input Video(avi)
if the greatest match for present pixel is the weight vector,
then the weight vectors that are updated according to are
weight vectors that belong in portion to the model of current
image pixel.
Color
conversion
3.3.5 Shadow Detection: The simple idea is that a cast
shadow blackens the background, whereas a moving entity
can darken it or not dependent on its color.
Cap model
Lighting
changes
compensation
Update
mixture
model
Calculate frame
probability
3.4 Dirichlet Process Gaussian Mixture
Model (DP-GMM): This method is a non-parametric
Bayesian method that spontaneously estimates the Number of
mixture components is automatically to model the pixels
background color distribution, e.g. Single mode pixel generate
at the trunk and in the sky, when the tree is waving forward
and backward in front of sky creates two modes pixels in the
area where braches wave, i.e. pixel transition among leaf and
sky regularly. If it requires more modes to denote multiple
leaf color this will take place automatically, and, of excessive
significance for term surveillance, this model will update with
time. However, two issues of standard DP-GMM model:
1) Update techniques of the existing model cannot cope with
the scene changes common in real-world applications;
2) if we used this model for continues video then more
computation and memory is required.
This model usages a Gaussian mixture model (GMM) for a
Per-pixel density estimation and followed by connected
component of regulation. Its mixture model has two
components foreground and background. It classifies values
based on their mixture components, which is allocated to the
foreground or the background. larger components belongs to
background and remaining belongs to foreground .
3.4.1 Block diagram: Video this is in AVI format is
taken because processing take place on AVI format. Then this
video is converted into frames and then color conversion is
take place that convert the color (RGB) images into gray
form. And reduces lighting effect occurs at the time of
capturing of video.The proposed method normally splits into
two parts—a per-pixel background model and a regularisation
step
3.4.1.1 Per-Pixel Background Model: every single
pixel has its multi-model density estimate, used to model
P(x/bg) where x is the pixel color channels vector. It can be
observed as the Dirichlet distribution prolonged to an infinite
number of components, which permits it to obtain the true
number of mixtures essential to indicate the data. Dirichlet
process, first using the stick breaking construction then
secondly using the Chinese restaurant process (CRP). Gives
clean description of concept is provided by stick breaking,
whereas the Chinese restaurant process integrates out
inappropriate variables and offers the formulation we actually
solve.
Dirichlet process: In probability theory, are family of
stochastic method whose realization is probability
distribution. It is frequently used in Bayesian interference to
define the preceding knowledge nearly the distribution of
random variables, that is, how possibly it is that the random
variable are distributed allowing to one or another specific
distribution.
Regulization
Object detected
Thresholding
Figure 3.4.1 : Block diagram of the approach
The Dirichlet process is presented by a base distribution H
and a positive real number α called the concentration
parameter. The base distribution is the probable the process
value, that is, the dirichlet process does distribution “around”
the base distribution in the way that a normal distribution
draws real numbers around it. If the base dtribution is constant
the distributions drawn from the Dirichlet process remain
surely discrete. The concentration parameter identifies how
robust this discretization: is in the limit of α→0, the
realization are all concentrated on a particular value, whereas
in the limit of α→∞ the realization becomes between the two
extreme the realization are discrete distribution with fewer
and fewer concentration as increases.
The Dirichlet process that may also be understood as the
infinite-dimensional generalization of the Dirichlet
distribution. In the related manner as the dirichlet distribution
is the infinite conjugate prior., In the same manner as the
Dirichlet distribution is the categorical distribution conjugate
prior, the dirichlet process is the infinite conjugate prior, nonparametric discrete distribution, A particularly important
application of the dirichlet process is the prior probability
distribution in infinite mixture model. Assume that the
production of values 𝑋1 ,𝑋2 ,… can be described with the
following algorithm.
Input: H (Probability distribution called Base distribution), α
(Positive real number called Concentration parameter)
1.
2.
Draw 𝑋1 from the distribution H.
For: n > 0
𝛼
1. With probability
draw X from H.
𝛼+𝑛−1
𝑛
𝑥
2. With probability
set 𝑥𝑛 = 𝑥, where 𝑛𝑥 is
𝛼+𝑛−1
the number of previous observations
𝑋𝑗 , 𝑗 < 𝑛
such that 𝑋𝑗 = 𝑋.
The 𝑋1 ,𝑋2 ,… the observations are dependent, then we have
to think through the previous results when producing value.
They are still, replaceable. This fact may be displayed by
computing joint probability distribution of the observations
and seeing that the resultant formula only based on which
value X occur among the observations and how many
duplications they each have. Procedure of the above
algorithm:
1.
1. Draw a distribution P from DP(H,𝛼)
2. Draw observations independently 𝑋1 , 𝑋2 ,… f rom P.
The Chinese restaurant process: The "Chinese
restaurant process" name is stated from the following analogy:
imagine an infinitely big restaurant having an infinite tables,
and capable to serve an infinite dishes. The restaurant in
question works with a slightly unusual seating policy whereby
new dinners are seated either at presently working table with
probability proportional to the number of guests at present
seated there, or at an unfilled table by means of probability
proportional to a constant. Guests who sit at engaged table
essential order the identical dish as those presently seated,
whereas guests assigned a new table are served a different
dish at random.The dishes distribution after J guests are
served is sample drawn as described below. Suppose that J
J
samples, , {θj } j = 1 samples must before been got according
to Chinese restaurant process, the (J + 1)th sample should be
drawn from
θ(J+1)~
𝟏
(H(S)+J)
𝑱
(H+ ∑𝒋=𝟏 𝜹𝜽𝒋 )
Where 𝛿𝜃 is atomic distribution . Understanding this, two
properties are clear:
1. Even if S is uncountable set, there is finite (i.e. non zero)
probability that two samples will have nearly the similar
value. A dirichlet process samples are discrete.
2. The dirichlet process shows a self-reinforcing property the
further every so often a identified value has been sampled in
the previous, the best probable it is to be sampled again
2.
Where 𝛽𝑘 are independent random variables with
the beta distribution Beta(1,𝛼). The correspondence
to ‘stick-breaking’ can be realized through seeing as
𝛽𝑘 the length of a part of a stick. We begin with a
unit-length stick and every step we halt a portion of
the remaining stick according to 𝛽′ 𝑘 and assign this
broken-off piece to 𝛽𝑘 . The formula can be
understood by observing that subsequently the first
k − 1 values have their portions allocated, the length
′
of the rest of the stick is∏𝑘−1
𝑖=1 (1 − 𝛽 𝑖 )and this
′
portion is broken according to 𝛽 𝑘 and becomes
assigned to 𝛽𝑘 .
The smaller α is, the fewer of the stick will be left
consequent values (on average), and resulting
further concentrated distributions.
In Stick breaking, stick is continuously break infinite times
and divides the samples into different Chinese restaurant sets.
Integrating out the draw from the DP indications to better
convergence, but more significantly replaces the infinite set of
stick with the computationally controllable finite set of tables.
3.4.1.2 Probabilistic Regularisation: Per-pixel
background model does not take information from the
adjacent or neighboring pixel so causes it susceptible to noise
and camouflage. Additionally, Gibbs sampling introduces
certain amount of noise i.e. dithering effect at the boundary
between foreground and background. This issues is resolved
by using markow random field, with node of each pixel
connected to four way neighborhoods. It is a binary labeling
problem where every single pixel corresponds to the either
foreground or to background.
Threshold value is set to 45, the pixel above threshold is
considered as foreground and below background. Cap model
is used to update the model Instead of repeating the same
calculation. Whatever output is obtained that is updated so
when next same output is generated then the previous outputs
are taken that are stored so reduces same calculations.
Stick breaking: A third approach to the Dirichlet process
is so-called stick-breaking process view, a dirichlet process
are distributions over a set S. As the distribution drawn is
discrete with probability 1.In the sticking-breaking procedure
opinion, we clearly use the discreteness and provide the
probability mass function of this random) discrete distribution
as:
𝒇(𝜽) = ∑ 𝜷𝒌 . 𝜹𝜽𝒌 (𝜽)
𝒌=𝟏
Where 𝛿𝜃𝑘 is indicator function which estimates to zero all
over, excepts for. δθk (θk ) = 1. Then this distribution is itself,
its mass function is parameterized through two of random
variable: the locations {𝜃𝑘 } ∞
𝑘=1 and the corresponding
probabilities{𝛽𝑘 }∞
𝑘=1 . In the current deprived of proof what
these random variables are.
The locations 𝜃𝑘 are identically and independent distributed
according to H, base distribution of the dirichlet process. The
probabilities 𝛽𝑘 are specified by a procedure approximating
the breaking of a unit-length stick:
′
𝛽𝑘 = 𝛽′ 𝑘 .∏𝑘−1
𝑖=1 (1 − 𝛽𝑖 )
3. HARDWARE:
This method is implemented in
FPGA.
3.1 FPGA:
FPGAs contain an array of programmable
logic blocks, and reconfigurable interconnects the logic blocks
together, like various logic gates that can be present inter
wired in different configuration. Logic blocks are configured
to achieve difficult combinational functions or simple logic
gates such as the AND or XOR.
FPGA receives the current and reference frame data
through serial communication bus UART. and stored data in
RAM and Perform the and generates the output. This output is
given to matlab through UART3.
3.2 RAM: A typical RAM cell has only four connections:
Data in (the D pin on the D flip-flop), data out (the Q pin on
the D flip-flop), Write Enable (often abbreviated WE; The C
pin on the D flip-flop), and Output Enable (the Enable pin
which we added).For our board, ram is 16 Mb. Ram stored
current frame, reference frame, intermediately generated
result (processing) and output.
3.3
UART:
The
universal
asynchronous
receiver/transmitter (UART) receives bytes of data and sends
the individual bits in sequential manner.at the destination, a
second UART re-assembles the bits into whole bytes Each
UART holds a shift register which convert serial data into
parallel form and vice versa. Transmission of single bit of bye
in single wire is less costly than transmission of multiple data
in parallel form in multiple wires.
The MSE, also denote to the error signal ei,= xi − yi, which
is the difference of the original and contaminated signals.
5.1.2 PSNR: It is the ratio of the maximum possible signal
power to the corrupting noise power that distracts the fidelity
of the representation. PSNR is usually stated with the help of
the logarithmic decibel scale.
The PSNR (in dB) is defined as:
4. OVERALL SYSTEM FLOWCHART:
𝑀𝐴𝑋 2
PSNR=10log10 (
𝑀𝑆𝐸
)
=20.log10 (
Input video
𝑀𝐴𝑋1
√𝑀𝑆𝐸
)
=20.log10 (𝑀𝐴𝑋1 )-20.log10 (𝑀𝑆𝐸)
(Matlab)
Frames to binary
image (Matlab)
Here, MAXI is the extreme possible pixel image value. When
the pixel is denoted with the help of bits per sample, this value
is 255. Mostly, when denoted using linear PCM with B bits
per sample, MAXI is 2B−1.
8 bit binary
image(FPGA)
5.1.3 Entropy: Entropy is a statistical degree of
Processsing
uncertainty that can be used to describe the texture of the
input image. Entropy is an index to evaluate the how much
information (quantity) contained in an image. Entropy is
defined as
E=-∑𝐿−1
𝐼=0 𝑝𝑖 log 2 𝑝𝑖
(FPGA)
Generates output
(FPGA)
Output 8 bit
stream
(Matlab)
Display output
(Matlab GUI)
Figure 4.1 Overall system flowchart
Where L is the total grey levels, 𝑝 = {𝑝0, 𝑝1, … . . 𝑝𝐿−1 } is
the probability distribution of each level
5.1.4 Correlation: Normalized cross correlation are used
to find out likenesses between current and reference image is
given by the following equation
NCC=
𝑛
∑𝑚
𝑖=1 ∑𝑗=1(𝐴𝑖𝑗∗ 𝐵𝑖𝑗 )
𝑛
2
∑𝑚
𝑖=1 ∑𝑗=1(𝐴𝑖𝑗 )
5.2 Results:
Output of different motion detection a
technique.
5.2.1
Backgound
Subtraction:
between current and reference frame
Ddifference
5. PERFORMANCE AND RESULTS:
5.1 Output Parameters:
5.1.1 MSE: MSE as a measure signal fidelity. The signal
fidelity compare the quantitative score of two signal so we can
described how to signals are similar and level of
distortion (noise) between them. Typically, it is considered
that one signal is an original signal, while the other signal is
distorted or contaminated with errors.
Figure 5.2.1: Output of background subtraction method
5.2.2 Frame difference: difference between two
consecutive frames
Assume that x = {xi|i = 1, 2, · · · , N} and y = {yi|i = 1, 2, ·
· · , N} are two finite-length, discrete signals (e.g., visual
images), where N is the number of signal samples (pixels, if
the signals are images) and xi and yi are the i th samples in x
and y, respectively.
The MSE is defined as:
1
2
MSE(x,y)= ∑𝑁
𝑖=1(𝑥𝑖− 𝑦𝑖 )
𝑁
Figure 5.2.2: Output of Frame Difference method
5.2.3 Neural Map method: pixel mapped into 3*3
mapping of current and reference frame.




Object detection
Video surveillance
Object tracking
Traffic monitoring
9. FUTURE SCOPE:
Figure 5.2.3: Output of Neural Map method
5.2.4 Background Subtraction with Dirichlet
process Gaussian mixture model (DP-GMM)
Combining information from
adjacent pixel in regulization does not fully achieve the
information available. A more challenging method of spatial
information transmission would be desirable-a conditional
Dirichlet process might offer this. Rapid complex lighting
variations are not controlled by this method i.e. it fails to
handle certain indoor lighting variations. Still, a more
sophisticated typical of the foreground and a clear model of
left object could further improve our method.
REFERENCES:
[1] Weiming Hu, Tieniu Tan,”A Survey on Visual
Surveillance of Object Motion and Behaviors” ieee
transaction on systems, and cybernetics-Part C:application
and revievs,vol . 34, no. 3,august 2004
[2] Qi Zang and Reinhard Klette,”Object Classification and
Tracking in Video Surveillance”,unpublished
Figure 5.2.4: Output of DP-GMM method
6. COMPARISION OF DIFFERENT
METHODS PARAMETERS
Method
/parameter
Backgroud
subtraction
Frame
Difference
Neural
Map
DPGMM
[3] Kinjal A Joshi, Darshak G. Thakore,”A Survey on Moving
Object Detection and Tracking in Video Surveillance
System”,International Journal of Soft Computing and
Engineering (IJSCE) ISSN: 2231-2307, Volume-2, Issue-3,
July 2012
[4] Zhen Tang, Zhenjiang Miao ,”Fast Background
Subtraction and Shadow Elimination Using Improved
Gaussian Mixture Model”, HAVE 2007 - IEEE International
Workshop on Haptic Audio Visual Environments and their
Applications ,Ottawa - Canada, 12-14 October 2007
[5] Nishu Singla ,”Motion Detection Based on Frame
Difference Method” International Journal of Information &
Computation Technology. ISSN 0974-2239 Volume 4,
Number 15 (2014), pp. 1559-1565
MSE
0.0115
0.0487
0.0245
0.0102
PSNR
67.5392
61.2563
64.243
68.673
[6] M. Julius Hossain, M. Ali Akber Dewan, and Oksam Chae
,”Edge Segment-Based Automatic Video Surveillance”,
Journal on Advances in Signal Processing Volume 2008,
Entropy
0.2950
0.0917
0.3766
0.2870
Correlation
0.8832
0.2783
0.8041
0.8812
[7] Tom S.F. Haines and Tao Xiang ,”Background
Subtraction with Dirichlet Process Mixture Models” ieee
transaction on pattern analysis and machine intelligence,vol.
36,no 4,April 2014
7. CONCLUSION:
Dirichlet method is suitable for
background modeling, and computationally scalable. This
method handles the dynamic background. Infinite no of
mixture components are used so whatever object is detected is
more accurate than the other methods. And also handles the
scene changes. It handles camouflage and shadow effect.
Other methods require more computation and give less
efficiency. From the above table, we can conclude that the
output or human body detection done by Background
subtraction with Dirichlet Process Gaussian Mixture model
method is more accurate than the another methods
8. APPLICATIONS:
[8] Lucia Maddalena and Alfredo Petrosino,,” A SelfOrganizing Approach to Background Subtraction for Visual
Surveillance Applications”, ieee transaction on image
processing vol. 17, No. 7, July 2008
[9]Babak Shahbaba and Radford Neal ,“Nonlinear Models
Using Dirichlet Process Mixtures”,Journal of Machine
Learning Research 10 (2009) 1829-1850
[10]Larissa ValmyAnd Jean Vaillant,”Bayesian Inference on
a Cox Process Associated with a Dirichlet Process”,
International Journal of Computer Applications (0975 8887)
Volume 95 - No. 18, June 2014
[11]Ibrahim Saygin Topkaya, Hakan Erdogan and Fatih
Porikli” Counting People by Clustering Person Detector
Outputs”, 2014 11th IEEE International Conference on
Advanced Video and Signal Based Surveillance (AVSS)