View
0
Download
0
Category
Preview:
Citation preview
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
32
SCHEMATIC MODEL FOR ANALYZING MOBILITY AND
DETECTION OF MULTIPLE OBJECT ON TRAFFIC SCENE
Pushpa D.1, Dr. H.S. Sheshadri2
1Asst. Prof: Dept of Information Science & Engg, Maharaja Institute of Technology,
Mysore, India 2Prof.: Dept. of Electronics and communication Engg, PES College of Engg,
Mandya, India
ABSTRACT
Detection, counting as well as gathering features to perform analysis of behavior
of natural scene is one of the complex processes to be design. My previous work has
introduced a simple model for identifying and counting the moving objects using simple
low level features. The current work is an extension of the previous work where the focus
is not only to detect and count the multiple moving objects but also to understand the
crowd behavior as well as exponentially reduce the issues of inter-object occlusion. The
image frame sequence is considered as input for the proposed schematic model.
Unscented Kalman filter is used for understanding the behavior of the scene as well as for
increasing the detection accuracy and reducing the false positives. Designed on Matlab
environment, the result shows highly accurate detection rate.
Keywords: Background Image, Foreground Image, Kalman Filter, Multiple Object
Detection, Object tracking, Object counting.
I. INTRODUCTION
The area of computer vision is still in its infancy. It has many real-world
applications, and many breakthroughs are yet to be made. Most of the companies in
existence today that have products based on computer vision can be divided into three
main categories: auto manufacturing, computer circuit manufacturing, and face
INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING
& TECHNOLOGY (IJCET)
ISSN 0976 – 6367(Print) ISSN 0976 – 6375(Online) Volume 4, Issue 3, May-June (2013), pp. 32-49 © IAEME: www.iaeme.com/ijcet.asp Journal Impact Factor (2013): 6.1302 (Calculated by GISI) www.jifactor.com
IJCET
© I A E M E
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
33
recognition. There are other smaller categories of this field that are beginning to be
developed in industry such as pharmaceutical manufacturing applications and traffic
control. Auto manufacturing employs computer vision through the use of robots that put
the cars together. Computer circuit manufacturers use computer vision to visually check
circuits in a production line against a working template of that circuit. Computer vision is
used as quality control in this case. The third most common application of computer
vision is in face recognition. This field has become popular in the last few years with the
advent of more sophisticated and accurate methods of facial recognition. Traffic control is
also of interest because computer vision software systems can be applied to already
existing hardware in this field. By traffic control, we mean the regulation or overview of
motor traffic by means of the already existing and functioning array of police monitoring
equipment. Cameras are already present at busy intersections, highways, and other
junctions for the purposes of regulating traffic, spotting problems, and enforcing laws
such as running red lights. Computer vision could be used to make all of these tasks
automatic. In the last few years, object detection and localization have become very
popular areas of research in computer vision. Although most of the categorization
algorithms tend to use local features, there is much more variety on the classification
methods. In some cases, the generative models show significant robustness with respect
to partial occlusion and viewpoint changes and can tolerate considerable intra-class
variation of object appearance [1, 2, 3]. However, if object classes share a high visual
similarity then the generative models tend to produce a significant number of false
positives. On the other hand, the discriminative models permit us to construct flexible
decision boundaries, resulting in classification performance often superior to those
obtained by only generative models [4, 5]. However, they contain no localization
component and require accurate localization in positions and scale. In the literature, the
standard solution to this problem is to perform an exhaustive search over all position and
scales. However, this exhaustive search imposes two main constraints. One of them is the
detector’s computational complexity. It requires large computational time for relatively
large number of objects. The second is the detector’s discriminance, since a large number
of potential false positives need to be excluded.
Problems like inter-object occlusion, different object moving at all different speed,
walking in arbitrary direction etc. will pose a huge challenge in the development as well
as design of multiple object detection as well as tracking from the crowded scene.
Majority of the video monitoring techniques are normally experimented with permanent
image capturing device which is comparatively unproblematic in terms of real time
application in order to accomplish image capturing parameters. Therefore, the current
research work has been experimented with monocular adjusted image capturing device,
which facilitates to design a distinctive connection between the subjects walking in real
time and image plane. The phenomenon is considered for isometric view of the subjects
walking. In Section 2, we review recent work on multi-object detection. In Section 3,
problem description is discussed in brief followed by proposed system in section 4.
Section 5 discusses about algorithm design that is implemented in the proposed system
which is followed by Section 6 that discusses about result discussion. Finally concludes
by summarizing the entire research work in section 7.
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
34
II. RELATED WORK
Moving object detection is the basic step for further analysis of video. Every
tracking method requires an object detection mechanism either in every frame or when
the object first appears in the video. It handles segmentation of moving objects from
stationary background objects. This focuses on higher level processing .It also decreases
computation time. Due to environmental conditions like illumination changes, shadow
object segmentation becomes difficult and significant problem. A common approach for
object detection is to use information in a single frame. However, some object detection
methods make use of the temporal information computed from a sequence of frames to
reduce the number of false detections. This temporal information is usually in the form of
frame differencing, which highlights regions that changes dynamically in consecutive
frames. A brief survey on the past research work is described here. One of the
significance work carried out by Koller et al. [6][7] is design of an algorithm which uses
offline calibrated camera step to aid the recovery of the 3D images, and it is also passed
through Kalman Filter to update estimates like location and position of the object. The
concept of Bayesian technique for image segmentation based on feature distribution is
also deployed where a statistical mixture model for probabilistic grouping of distributed
data is [8] adopted. It is mainly used for unsupervised segmentation of textured images
based on local distributions of Gabor coefficients. Toufiq P. et al., in [9] describes
background subtraction as the widely used paradigm for detection of moving objects in
videos taken from static camera which has a very wide range of applications.
Polana and Nelson [10] has adopted spatial-temporal discretization which is
allocated to every region of image, accomplished normally by overlapping a grid on the
image and its mean visual stream. The episodic movement prototype is estimated by
performing Fourier analysis on the resulting attribute vector. The use of movement field
by Shio and Sklansky [11] has shown the accomplishment by correlation technique over
consecutive image frames, which results in identification of movement regions with
uniform direction when smoothening is done along with quantization of image regions.
Deployment of clusters of pixels has been done by Heisele et al. [12] to be attained
by different grouping methods as fundamental constituents for tracking. The grouping
analysis is performed by both spatial data and color information. The concept behind this
is that appending spatial data generates grouping more steady than deploying only color
data. There has been also abundant use of k-means algorithm while working on clustering
techniques as evident in Cutler and Davis [13] for designing a change detection
algorithm. The prototype also facilitates a map of the image elements signifying moving
subjects along with precise thresholding.
Use of wavelets transform has been seen in the work of Papageorgiou and Poggio
[14] for characterizing a moving object shape and then examines the transform technique
of the image to evaluate the movement patterns. Implementation of two dimensional
contour shape analyses is seen in the work of Wren et al. [15] which tends to recognize
different parts of body of the moving subjects using heuristics techniques. Cai and
Aggarwal [16] proposed a model with a uncomplicated head-trunk replica to trail humans
across multiple image capturing devices. Beymer and Konolige [17] have worked for
fitting a trouble-free shape on the candidate image.
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
35
Dimitrios Makris [18] with the non-trivial problem of performance evaluation of
motion tracking. We propose a rich set of metrics to assess different aspects of
performance of motion tracking. We use six different video sequences that represent a
variety of challenges to illustrate the practical value of the proposed metrics by evaluating
and comparing two motion tracking algorithms. Max e.t. al [19] describes performance
results from a real-time system for detecting, localizing, and tracking pedestrians from a
moving vehicle. It achieves results comparable with alternative approaches with other
sensors, but offers the potential for long-term scalability to higher spatial resolution,
smaller size, and lower cost than other sensors. But issue of this work is it cannot segment
people or objects in close contact. Jifeng [20] has presented a novel object tracking
algorithm where the author has used the joint color texture histogram to represent a target
and then applying it to the mean shift framework.
To analyze images and extract high level information, image enhancement, motion
detection, object tracking and behavior understanding researches have been studied. In
this section, we have studied and presented different methods of moving object detection
from prior research work. Various methods like background subtraction, temporal
differencing, statistical methods were explored. Detection techniques into various
categories, here, we also discuss the related issues, to the moving object detection
technique. The drawback of temporal differencing is that it fails to extract all relevant
pixels of a foreground object especially when the object has uniform texture or moves
slowly. When a foreground object stops moving, temporal differencing method fails in
detecting a change between consecutive frames and loses the track of the object. The
survey also identified background subtraction method in deep because of its
computational effectiveness and accuracy. This section gives valuable insight into this
important research topic and encourages the new research in the area of moving object
detection as well as in the field of computer vision. It was also found that research on
object tracking can be classified as point tracking, kernel tracking and contour tracking
according to the representation method of a target object. In point tracking approach,
statistical filtering method has been used to estimating the state of target object. Kalman
filter and particle filter are the most popular filtering method. In kernel tracking approach,
various estimating methods are used to find corresponding region to target object. Now a
day, the most preferred and popular kernel tracking techniques are based on mean-shift
tracking and particle filter. Contour tracking can be divided into state space method and
energy function minimization method according to the way of evolving of contours. The
survey has witnessed various publications that are focused on addressing issues on object
detection. Finally, Table 1 will show the evidence of some recent publications related to
the issue of Object detection. Majority of the Journals based on moving objects are
selected for the review purpose as shown below:
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
36
Table 1 Existing Approaches used in Moving Object Detection
Year Authors Problem Focused Techniques Used Remark
2010
Djamel, Nicolas [21] Counting people Skelton graph,
Background subtraction
Objects other than human
are not considered
Conte e.t. al [22] Counting people Motion vector,
Support vector Regressor
Objects other than human
are not considered
Burkert e.t. al [23] Monitoring people based
on aerial images
Microscopic
& Mesoscopic parameter,
trajectory
Inter-object Occlusion not
focused on
Conte e.t. al [24] Counting Moving
People
Detection using interesting
points
Objects other than human
are not considered
Widhalm [25] Flow of Pedestrian Growing Neural Gas,
Dynamic Time Warping
Objects other than human
are not considered, Inter-
object Occlusion not
focused on
2011
Dehghan e.t. al [26] Automatic Detection &
Tracking of Pedestrians
Static Floor Field,
Boundary Floor Field,
Dynamic Floor Field
Addressed Occlusion, good
results obtained, not
focused on multiple objects
moving
Spampinato [27] Event Detection in
crowd
Lagrangian Particle
Dynamics
Results found with Issue in
noises in frames affecting
the performance of flow
map estimation
Jae Kyu Suhr [28] Moving Object
Detection
Background compensation
method using 1-D feature
matching & outlier
rejection
Not focused on inter-object
occlusion, Multiple objects
are not considered
Lei, Su, Peng [29] Maritime
Surveillance
Trajectory Pattern Mining Better result but cannot be
used in crowded scenes
Sirmacek [30] Tracking people from
airborne images
Kalman filter Not focused on inter-object
occlusion, Multiple objects
are not considered
Year Authors Problem Focused Techniques Used Remarks
2012
Cui, Zheng [31] Moving Object
Detection
Graph cuts Not focused on inter-object
occlusion, Multiple objects
are not considered
Chen e.t. al [32] Counting People in
crowd
2-Stage Segmentation Better result, but impact of
occlusion & multiple
objects are not considered
Gowsikhaa [33] Human Activity
Detection
Artificial Neural Network Topic is narrowed down to
face recognition system,
Occlusion issue is not
addressed
Rosswog [34] Object detection in
crowd
SVM Occlusion issue is not
addressed
Arif [35] Counting Moving Object Neural Network Illumination state &
Occlusion issues are not
addressed
Although, we can’t conclude that the previous work has got any demerits or
disadvantages, but it can be seen that with each discoveries of research, there is a new idea
which keeps the research on ongoing pace. But, in nutshell, it can be said that majority of the
work has not much integrated the process of Multiple (different) objects detection along with
solution for inter-object occlusion and reduction of false positives. Hence a research gap can
be explored in this viewpoint.
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
37
III. PROBLEM DESCRIPTION
Searching for a known object in a specified scene and locating a given object are
inherently different problems. Object recognition is a computationally expensive process.
Fast algorithms are essential at all stages of the recognition process. Object recognition is
tricky because a combination of factors must be considered to identify objects. These factors
may include limitations on allowable shapes, the semantics of the scene context, and the
information present in the image itself. Given a video clip, the initial problem is segregating
it into number of frames. Each frame is then considered as an independent image, which is in
RGB format and is converted into Gray scale image. Next the difference between the frames
at certain intervals is computed. This interval can be decided based on the motion of moving
object in a video sequence. If the object is moving quite fast, then the difference between
every successive frame has to be considered, posing a difficulty in designing smart object
identification system.
The key issue of multiple targets tracking in the image sequences is the occlusion
problem. During the occlusion time, unpredictable events can be activated. Thus, the tracking
failure and missed tracking often happened. To cope with this problem, a behavior analysis of
multiple targets is required the first time. A specific target enters into the specific field of
view, and then it is moving, stopping, interacting with other targets. And finally, it leaves the
field of view. We can classify the behaviors of multiple targets accordingly as
1. A specific target enters into the scene.
2. Multiple targets enter into the scene.
3. A specific target is moving and forms a group with other targets, or just moves beside
other targets or obstacles.
4. A specific target within the group leaves a group.
5. A specific target continues to move alone, or stops moving and then starts to move
again.
6. Multiple targets in a group continue to move and interact between them, or stop
interacting and then start to move again.
7. A specific target leaves a scene.
8. A specific group leaves a scene
The events of (1), (4), (5), and (7) can be tracked using general tracking algorithms.
However, the events of (2), (3), (6) and (8) cannot be tracked reliably. Thus, to resolve this
problem, we propose the occlusion reasoning method that detects occlusion activity status
using Kalman Filter.
IV. PROPOSED SYSTEM
The prime aim of the proposed work is to design a robust framework that is capable of
identifying mobile objects using sequences of frames captured from video. The research work
is designed not only for its utility in crowd detection that was presented in my previous work
[36] but now the proposed framework will be used for extensive analysis of the video dataset.
Various knowledge and data about the coordinates and characteristics of the mobile object at
various instance of time as the fundamental of identifying abnormal mobile object detection
pertaining to its movement. The flow of the proposed system is as highlighted in Fig.1.
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
38
Figure 1 Process Flow of the proposed system
The proposed study is designed to accept the sequential frames of the video, which is
converted to blob image or binary images, Image plane is defined for the purposed of frame
representation as well as for the purpose of estimating the centroid location.(we are using the
relative distance from centroid in image plane from the position of foreground object). A
Gaussian distribution is applied to understand the pixel density, so that distinction of every
pixel changes is easy to find/estimate. Another purpose of Gaussian distribution is to find the
peak in pixel intensity distribution which will be used for threshold estimation. The white
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
39
pixels are compared with the threshold in order to distinguish the foreground and background
object. However, after estimating the relative position of newly detected foreground object,
background is also found and finally state transition matrix (STm) is compared with the
threshold value (Tm). The condition is checked to predict distinctly about the foreground and
background object. The Kalman filter is the best part of the contribution.
The proposed system uses the frames captured from the still camera, where the
visibility is restricted to a defined area of focus of the aperture of the camera. This method is
adopted to analyze the preciseness of the mobility of the object, which could not be done on
mobile video sensors. The system uses distinctive technique to filter out the foreground
objects from the background for better accuracy. Basically, the background model is
considered as prime module as the experimental framework designs for every pixels of the
captured frame as a distribution of the typical values considered by background module. The
main concept behind this is the most frequent values of a pixel that are corresponding to the
background metaphors. The system is designed in such a way that the iterative mobility in the
background is also captured. That will mean that even a slightest displacement on any object
on the considered visibility screen will also be captured.
Figure 2 showing Original Scene, Mask
Figure 3 showing Original Scene, Mask
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
40
The categorization of the pixel into background or foreground is designed
depending on the collected module of background using outlier-detection. According to
this design, the pixels with the intensity values sufficiently distinctive from the
background distribution are indexed as foreground. The framework will also use gradual
alteration in the illumination factor by spontaneously upgrading the background model of
every pixel in the frames. The same phenomenon will also be applicable to the static
objects in the considered scene. An example input image and the corresponding pixel
labeling computed from a background model can be seen on the left where foreground
pixels were labeled as white. The person crossing the street can be clearly seen in Fig.2
and Fig.3. From Fig.2 and Fig.3, by lumping large connected foreground regions into
blobs, foreground objects, such as the person seen crossing the street is identified.
However, the left images also show that an additional foreground object was erroneously
detected in the top right corner due to violent motion of the trees in the background. Such
occasional misdetections are taken into account at later stages of the object detection
algorithm.
The consecutive stage will include a sequence of the object blobs together across
various frames of the scene. In order to accomplish this task, various attributes like
position, velocity, and size are used to represent every object block. Unscented Kalman
filter is specifically used in this work for this purpose for determining the actual
coordinates of the mobile object identified on the scene-frames. The main purpose of
adopting this step is to truncate the possibility of noise from blob estimations which could
possibly lead to generation of the false-positives values.
V. ALGORITHM DESIGN
The proposed work is performed on 32-bit Windows OS. In order to smoothly run
the heavy computation from large datasets of frames, it is highly essential to use high
configuration system with min 2.20 GHz processor with Intel core-i3 along with 4 GB
RAM size. The programming is done on Matlab environment due to its faster
computation of complex problems. The image sequence datasets has been taken from
[36].
By deploying the computed indexing, foreground objects are identified, however,
direct use of these blobs for tracking is sensitive to noise and fails to determine object
identities at different times in a sequence. In order to couple a sequence of state
measurements, that is, relevant blob properties such as the centroid, centroid velocity
between consecutive frames, or size into a denoised state trajectory, The filter operates as
a two-stage process switching between absorbing evidence from the most recent
foreground blob and forming a predictive distribution for the next. In the context of
tracking multiple objects simultaneously, the key difficulty lies in determining which
Unscented Kalman filter should absorb which new state measurement. If this association
task is not solved adequately then Unscented Kalman filters are corrupted with wrong
information.
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
41
Algorithm: Detection of object mobility
Input: Image sequence datasets
Output: Appropriate detection of object movement in visualization
START 1 Load set of image sequence
2 Returns data matrix
3 Reading Indexed individual frames
4 Reading RGB components of the scene
5 Reading frame number
6 Estimate the log of the Gaussian distribution for each point in x using µ and σ.
7 x=matrix of column-wise data points.
8 µ =column vector
9 σ =Covariance matrix of respective size
10 Estimate the log probability for each column in x-centered.
11 Deploy other parameters:
12 weight=matrix of mixing proportion
13 σ =Covariance
14 Active_Gaussian=components that were either matched or replaced.
15 Background_Threshold=Threshold standard deviation for outlier detection on
each Gaussian.
16 K= Quantity of components used in the mixtures.
17 Alpha=How fast for adapting the component weights
18 Rho=How fast to adapt the component means and covariances
19 Threshold_Deviation=Threshold used for finding matching versus unmatched
components
20 INIT_Var= Initial variance for newly placed components
21 INIT_MixProp= Initial prior for newly placed components
22 COMPONENT_THRESH: Filter out connected components of smaller size
23 BACKGROUND_THRESH: Percentage of weight that must be accounted for
by background models
24 Design a helper function for structure of Kalman filtering
25 Define Kalman State type
26 Create best background image
27 Computes a β probability density on image coordinates pos
28 Set the state transition matrix
29 Update using position only
30 Update using position and velocity
31 Compute a cost matrix
32 Computes the squared distance between the centroid predictions of a list of objects
and a list of observed blob centroids.
33 Compute a cost matrix
34 Compute the square distance between Kalman filter predictions and
observations
35 Create foreground_map
36 Binary matrix
37 Define chromaticity coordinates
38 Show visualization of object detection
END
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
42
VI. RESULT DISCUSSION
In order to accomplish the research work, the whole detection protocol was assessed
on a number of different sequences in order to analyze its accurateness and weakness. The
previous work [36] done is able to detect objects from crowded scene, but false positive are
quite moderate in number. Hence, this paper will use the technique of foreground detection
that can be used for extracting useful information under various distinctive conditions,
particularly when the data stream is recoded in different color space. The cumulative results
obtained are as below in Fig.3
Figure 4 Detection window showing object detection
The above Fig. 4 shows frame rates and movement of each object, along with its
count. The interesting thing is that it also shows every minute movement even for shaking
leaves of tree. Manual inspection of various detection results suggested that multiple object
tracking using the devised association technique performs adequately in many situations.
Figure 5 Result showing successful detection rate
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
43
Figure 6 Result showing successful detection of multiple-object and counting
Figure 7 Result showing preciseness in multiple-object detection counting
The assessment is done considering three case scenarios e.g. i) Successful detection
rate of object: Fig.5 shows very effectively both truck and human along with the person
riding bicycle. ii) Successful detection of multiple-object and counting: Fig.6 shows very
distinctively the human as well as truck as a moving objects. iii) Preciseness in multiple-
object detection counting: Fig.7 shows an accurate detection of a person riding bicycle, a
person walking, and a person .walking along with pram. All the three objects are accurately
shown and counted too. The above implemented result shows better detection rate for
multiple object with precise bounding box over the object. It successfully counts the objects
too. The previous model has some detection issue due to inter-occlusion. However, 95% of
the inter-occlusion issue are overcome in this model with highly precise detection rate.
VII. PERFORMANCE DISCUSSION
For evaluating the performance of the application, the selected video clip in chosen to
understand the various frames as mentioned in previous section. The classifications of the
images are done and consecutively Machine is trained where the total number of classes has
to be defined with class name. Finally object recognition module is executed. This video
segmentation method was applied on three different video sequences one of which are
depicted in Fig 8 and Fig 9, which represents original pictures, mask, and detected Object.
Figure 8 showing Original Scene, Mask, Detected Object (First Sequence)
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
44
Figure 9 showing Original Scene, Mask, Detected Object (Second Sequence)
For the first video sequence Figure 2 & 3 shows the original image, the background
registered image, and image obtained after background subtracted, and finally shows the
count of the detected objects. The same is repeated for the next video sequence.
Figure 10 Object counting for moving objects
The system is able to track and count most vehicles successfully. Although the
accuracy of vehicle detection was 100%, the average accuracy of counting vehicles was 99%
(Fig 10). Each video sequences are tested with various types of object and also in complex
background too. The application is successfully able to identify as well as track various
pedestrians separately along with intersection. The application also successfully conducts
object recognition (Fig 11). This is due to noise which causes detected objects to become too
large or too small to be considered as a vehicle.
Figure 11 Moving Object recognition for separated objects moving
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
45
Figure 12 Different Moving object recognition in one scene.
However, two vehicles will persist to exist as a single vehicle if relative motion
between them is small and in such cases the count of vehicles becomes incorrect.
Figure 13 Performance Analysis for Accuracy
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
46
Figure 14 Performance Analysis for Object Identification
Figure 15 Performance Analysis for Object Recognition.
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
47
Fig 13 represents the performance analysis of accuracy level considering all the three
videos for evaluation. The accuracy rate, identification and recognition rate is better
approximated to 100% respectively. Fig 14 shows the identification analysis which represents
the better identification rate in all the AVI files. Fig 15 shows better and enhanced
performance in recognition for all types of moving object.
VIII. CONCLUSION
This paper has presented a novel and highly precise model of detecting multiple
objects in real time street scene with highly uncertainty of the types of objects, their
movements, and their count rates. The previous model presented has been addressed for
object detection from crowded scene with better detection rate, but even some moderate false
positives were also raised due to inter-occlusion and variances in illumination. However,
using unscented Kalman filter such issues were overcome in this model. The scope of the
application for this model is now much higher. If the previous model could be used to detect
multiple object than this model can be used to analyzed the behavior of object in the real time
scene. The simulation result shows highly precise detection rate. However, these will
eventuality limits the specification of the model in terms of segregating the targeted object
with not so important object. For example, due to proposed algorithm that has used both
foreground and background model, the framework detects even the minor movement of
leaves of trees. This property might be little unwanted when we want to design an application
specific to analyze pedestrian or car or some major moving object visually that keeps highest
importance in application. Hence, the next further work will be enhanced the model using
homographic transformation
REFERENCES
[1] S. Josef, C. Bryan, Russell, A. Alexei, A. Zisserman, and T. William, “Discovering
objects and their location in images,” Proc. ICCV’05, Beijing, China, pp.370–377, 2005.
[2] R. Fergus, P. Perona, and A. Zisserman, “Weakly supervised scale-invariant learning
of model for visual recognition,” IJCV, vol.71, no.3, pp.273–303, 2007.
[3] B. Leibe, A. Leonardis, and B. Schiele, “Combined object categorization and
segmentation with an implicit shape model,” Proc. ECCV’04, Prague, pp.17–32, 2004.
[4] A.Y. Ng, and M.I. Jordan, “On Discriminative vs. Generative Classifiers: A
comparison of logistic regression and naive Bayes,“ Proc. NIPS’01, Vancouver, British
Columbia, Canada, pp.841–848, 2001.
[5] V. Ferrari, L. Fevrier, F. Jurie, and C. Schmid, “Group of adjacent contour segment
for object detection,” PAMI, vol.30, no.1, pp.30–51, 2008.
[6] D. Koller, J. Weber, T. Haung and J. Malik, “Moving Object Recognition and
Classification based on Recursive Shape Parameter Estimation”, In Proceedings of the 12th
Israeli Conference on Artificial Intelligence, Computer Vision and Neural Networks, pp. 359-
368, Tel-Aviv, Israel, December, 1993.
[7] D. Koller, J. Weber, T. Haung, J. Malik, G. Ogasawara, B. Rao and S.Russel,
“Towards Robust Automatic Traffic Scene Analysis in Real-Time”, In Proceedings of the
12th International Conference on Pattern Recognition (ICPR-94), pp. 126-131, Jerusalem,
Israel, October 9-13, 1994.
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
48
[8] Puzicha J., Hofmann T. and Buhmann J. M., “Histogram Clustering for Unsupervised
Image Segmentation”, In IEEE Computer Society Conference Computer Vision and Pattern
Recognition, Fort Collins, CO, pp. 602- 608, June, 1999.
[9] Toufiq P., Ahmed Egammal and Anurag Mittal, “A Framework for Feature Selection
for Background Subtraction”, In Proceedings of IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR’06), 2006.
[10] Polana, R. and Nelson, R. C., Detection and recognition of periodic, nonrigid motion,
Int. J. Comput. Vision 23(3): 261–282., 1997
[11] Shio, A. and Sklansky, J. Segmentation of people in motion, IEEE Workshop on
Visual Motion, pp. 325–332., 1997
[12] Heisele, B., Kressel, U. and Ritter, W. Tracking non-rigid, moving objects based on
color cluster flow, Proc. of IEEE Conference on Computer Vision and Pattern Recognition,
San Juan, pp. 257–260, 1997
[13] Cutler, R. and Davis, L. S., Robust real-time periodic motion detection, analysis and
applications, IEEE Trans. Patt. An. Mach. Int. 22(8): 781–796., 2000
[14] Papageorgiou, C. and Poggio, T., A pattern classification approach to dynamical
object detection, ICCV ’99: Proceedings of the International Conference on Computer
Vision-Volume 2, IEEE Computer Society, Washington, DC, USA, p. 1223, 1999
[15] Wren, C., Azarbayejani, A., Darrell, T. and Pentland, A., Pfinder: Realtime tracking
of the human body, IEEE Trans. Pattern Anal.Mach.Intell. 19(7): 780–785., 1997
[16] Cai, Q. and Aggarwal, J., Tracking human motion using multiple cameras, Proc. of
International Conference on Pattern Recognition, Vienna, pp. 68–72., 1996
[17] Beymer, D. and Konolige, K., Real-time tracking of multiple people using continuous
detection., 1999
[18] Fei Yin, Dimitrios Makris, Sergio Velastin, Performance Evaluation of Object
Tracking Algorithms, In 10th IEEE International Workshop on Performance Evaluation of
Tracking and Surveillance (PETS2007), Rio de Janeiro, Brazil (October 2007)
[19] Max Bajracharya, Baback Moghaddam, Andrew Howard, Shane Brennan, Larry H.
Matthies, Results from a Realtime Stereo-based Pedestrian Detection System on a Moving
Vehicle, Proceedings of the IEEE ICRA 2009 Workshop on People Detection and Tracking
Kobe, Japan, May 2009
[20] Jifeng Ning, Lei Zhang, David Zhang, Chengke Wu, Robust object tracking using
joint color-texture histogram, International Journal of Pattern Recognition and Artificial
Intelligence Vol. 23, No. 7, 2009
[21] Djamel Merad, Kheir-Eddine Aziz, Nicolas Thome, Fast People Counting Using
Head Detection From Skeleton Graph, 2010 Seventh IEEE International Conference on
Advanced Video and Signal Based Surveillance.
[22] Donatello Conte, Pasquale Foggia, Gennaro Percannella, Francesco Tufano, A
method for counting people in crowded scenes, EURASIP Journal on Advanced Signal
Processing, 2010
[23] F. Burkert, F. Schmidt, M. Butenuth, S. Hinz, People tracking and trajectory
interpretation in aerial image sequences, IAPRS, 2010
[24] D. Conte, P. Foggia, G. Percannella, F. Tufano and M. Vento, Counting Moving
People in Videos by Salient Points Detection, International Conference on Pattern
Recognition, 2010
[25] Peter Widhalm, Norbert Brandle, Learning Major Pedestrian Flows in Crowded
Scenes, International Conference on Pattern Recognition, 2010
International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 3, May – June (2013), © IAEME
49
[26] Afshin Dehghan, Haroon Idrees, Amir Roshan Zamir and Mubarak Shah*, Keynote:
Automatic Detection and Tracking of Pedestrians in Videos with Various Crowd Densities,
Springer-Verlag Berlin Heidelberg 2011
[27] C. Spampinato, A. Faro, S. Palazzo, Event Detection in Crowds of People by
Integrating Chaos and Lagrangian Particle Dynamics, Proc. 3rd Int. Conf. on Information and
Multimedia Technology (ICIMT 2011), Dubai, UAE, December 28-30, 2011.
[28] Jae Kyu Suhr, Moving Object Detection for Static and Pan-Tilt-Zoom Cameras in
Intelligent Visual Surveillance, A Doctorial Thesis, Graduate School of Yonsei University,
2011
[29] Po-Ruey Lei, Ing-Jiunn Su , Wen-Chih Peng, Wei-Yu Han, and Chien-Ping Chang, A
Framework of Moving Behavior Modeling in the Maritime Surveillance, Journal of Chung
Cheng Institute of Technology, 2011
[30] Beril Sirmacek, Peter Reinartz, Kalman filter based feature analysis for tracking
people from airborne images, ISPRS Workshop High-Resolution Earth Imaging for
Geospatial Information, Hannover, Germany, June 2011.
[31] Yuyong Cui, Zhiyuan Zeng, Weihong Cui, Bitao Fu, Moving Object Detection Based
Frame Difference and Graph Cuts, Journal of Computational Information Systems, Vol 8,
2012
[32] Chao-Ho Chen, Tsong-Yi Chen, Da-Jinn Wang, and Tsang-Jie Chen, A Cost-E_ective
People-Counter for a Crowd of Moving People Based on Two-Stage Segmentation, Journal
of Information Hiding and Multimedia Signal Processing, ISSN 2073-4212, Volume 3,
Number 1, January 2012
[33] Gowsikhaa D., Manjunath, Abirami S., Suspicious Human Activity Detection from
Surveillance Videos, (IJIDCS) International Journal on Internet and Distributed Computing
Systems. Vol: 2 No: 2, 2012
[34] James Rosswog, Kanad Ghose, Detecting and Tracking Coordinated Groups in
Dense, Systematically Moving, Crowds, SIAM-2012
[35] Muhammad Arif, Muhammad Saqib, Saleh Basalamah and Asad Naeem, Counting of
Moving People in the Video using Neural Network System, Life Science Journal 2012
[36] Pushpa D., Dr. H.S. Sheshadri, Precise multiple object identification and tracking
using efficient visual attributes in dense crowded scene with regions of rational movement,
IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 2, No 2, March 20
[37] V Purandhar Reddy, “Object Tracking by DTCWT Feature Vectors”, International
Journal of Computer Engineering & Technology (IJCET), Volume 4, Issue 2, 2013,
pp. 73 - 78, ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375.
[38] Ravi Kumar Jatoth and Dr.T.Kishore Kumar, “Real Time Implementation of
Unscented Kalman Filter for Target Tracking”, International journal of Electronics and
Communication Engineering & Technology (IJECET), Volume 4, Issue 1, 2013,
pp. 208 - 215, ISSN Print: 0976- 6464, ISSN Online: 0976 –6472.
Recommended