CHAPATER 1

INTRODUCTION TO OBJECT TRACKING

Object tracking is an important task within the field of computer vision. The proliferation
of high-powered computers, the availability of high quality and inexpensive video
cameras, and the interesting need for automated video analysis has generated a great deal
of interest in object tracking.

In its simplest form, tracking can be defined as a method of following an object

through successive image frames to determine its relative movement with respect to other
objects. In other words, a tracker assigns consistent labels to the tracked objects in
different frames of video.

One can simplify tracking by imposing constraints on the motion or appearance of

objects. One can further constrain the object motion to be of constant velocity or
acceleration based on prior information. Prior knowledge about the number and the size
of objects, or the object appearance and shape can also be used to simplify the problem.

Numerous approaches for object tracking have been proposed. These primarily differ
from each other based on the way they approach the following question: which object
representation is suitable for tracking? Which image features should be used? How
should the appearance and shape of object be modelled? The answers to these questions
depend on the context/environment in which the tracking is performed. A large number of
tracking methods have been proposed which attempt to answer these questions for variety
of scenarios.
Figure 1 shows the schematic of a generic object tracking system. As can be seen, visual
input is usually achieved through digitized images obtained from a camera connected to a
digital computer. This camera can be either stationary or moving depending on the
application. Beyond image acquisition, the computer performs the necessary tracking and
any higher-level tasks using the tracking result.

Fig 1: schematic of a generic object tracking system. The camera obtains visual images,
and the computer tracks the seen objects.
 OBJECT DETECTION

Detecting regions that corresponds to moving objects in video sequence plays a very
important role in many computer vision applications.
In simplest form, Object detection from video sequence is the process of detecting the
moving objects in frame sequence using digital image processing techniques. Moving
object detection is the basis of moving object identification and tracking.

Challenges of moving object detection:

 Loss of information caused by the 3D world on a 2D image
 Noise in images
 Complex object motion
 Non-rigid or articulated nature of objects
 Partial or full object occlusions
 Complex object shapes
 Scene illumination changes

The survey present here covers object detection algorithms appeared in recent literature.
We present the taxonomy of object detection algorithms in which the algorithms are
classified into five major categories. The advantages and disadvantage of the algorithms
considered are tabulated.
 CHAPATER 2
LITERATURE SURVEY
2.1. Literature Review

The techniques stated in [3] ranges from very basic algorithm to state of the art published
techniques categorized based on speed, memory requirements and accuracy. They used
methods such as frame difference technique, Real time background subtraction and
shadow detection technique, adaptive background mixture model for real time tracking
technique. They used algorithms ranges from varying levels of accuracy and
computational complexity. Some of them can also deal with real time challenges like
snow, rain, moving branches, objects overlapping, light intensity or slow moving objects.
The problems of achieving high detection rate with low false alarm rate for human
detection and tracking in video sequence is to maximize performance and improve
response time. The stated causes are the effect of scene complexity, scale changes and
scene background-human interactions. A two-step processing solution which is, human
detection, and human tracking with two novel pattern classifiers presented in [4]. There
are three basic phases in video examination: detection of interesting objects in video
scene, tracking of such objects from frame to frame, and analysis of object tracks to
recognize their activities. Detecting humans from video is a challenging problem owing
to the motion of the subjects. In [6] they developed a detector for moving people in
videos with possibly moving cameras and backgrounds, testing several different coding
schemes of moving object and showing that orientated histograms of differential optical
flow give the maximum performance. Motion-based descriptors are combined with
Histogram of Oriented Gradient appearance descriptors. Achieved detector is tested on
several databases includes a challenging test set taken from video and containing wide
ranges of position, motion and background imbalance, including rotating cameras and
backgrounds. [14] In [7], they have analyzed moving object detection techniques, frame
difference and the approximate median method. The frame differentiating has been
adopted for the reference frame and the step length. They have suggested the moving
object detection and object tracking by using the modified frame difference method. In
the surveillance system for video captured by single camera is considered for the space
under the observation. This method is experiment on almost ten videos and the results are
quite satisfactory.
 CHAPATER
TECHNICAL ACTIVITIES

Working

The background subtraction method which is used for framing the moving object from its
background which requires a following steps:
a) Reference frame selection (RFS): In it the initial frame is selected as the reference
frame.
b) Step Length: Appropriate step length has been selected on the basis of experimental
results.
c) Removing Noise: Noise is affecting the accuracy and performance of system so it has
to be removed.
d) Moving object detection (MOD): to detect the moving object from the frame
difference with the help of background subtraction methods like Frame difference,
approximate median and Modified frame difference methods.
e) Suspicious Activity: The bounding box is constructed in the isolated area of interest
from video sequence and the object is tracked according to its movement.
f) Rise alert: After tracking object the recorded sound will be generate for alert. Below
image is the example of these above steps.
In [5] cascade-of-rejectors approach with the Histograms of Oriented Gradients features
to achieve a fast and accurate human detection system. The features used are Histograms
of Oriented Gradients of variable-size blocks that capture salient features of humans
automatically. Using algorithm for feature selection, it identifies the appropriate set of
blocks, from a large set of possible blocks. It uses the integral image representation and a
rejection cascade which significantly speed up the computation. For an image, the system
can process 5 to 30 frames per second depending on the density in which it scans the
image, while maintaining an accuracy level similar to existing methods. In [1], author
specified new algorithm for detecting moving objects from a static background scene to
detect moving object based on background subtraction.
 OBJECT REPRESENTATION

In a tracking scenario, an object can be defined as anything that is of interest for further
analysis. For instance, boats on the sea, fish inside an aquarium, vehicles on a road,
planes in the air etc are a set of objects that may be important to track in a specific
domain. Objects can be represented by their shapes.

In this section, we will describe the object shape representations commonly

employed for tracking.

1. Points:
The object is represented by a point, that is, centroid (fig 2(a)) or by a set of points
(fig 2(b)). The point representation is suitable for tracking objects that occupy small
regions in an image.

2. Primitive geometric shapes:

Object shape is represented by a rectangle, ellipse (fig 2©, (d)) etc. primitive
geometric shapes are more suitable for representing simple rigid objects, they are also
used for tracking non rigid objects.

3. Object silhouette and contour:

Contour representation defines the boundary of an object (fig 2(g), (h)). The
region inside the contour is called the silhouette of the object (fig 2(i)). Silhouette and
contour representations are suitable for tracking complex non rigid shapes.

4. Articulated shape models:

Articulated objects are composed of body parts that are held together with joints.
For example, the human body is an articulated object with legs, hands, head feet
connected by joints. In order to represent an articulated object, one can model the
constituent parts using cylinders or ellipses as shown in fig 2(e).
5. Skeletal models:
Object skeleton can be extracted by applying medial axis transform to the object
silhouette. This method is commonly used as a shape representation for recognizing
objects. Skeleton representation can be used to model both articulated and rigid objects
(fig 2(f)).

Object representations are usually chosen according to the application domain.

For tracking object, which appear very small in an image, point representation is usually
appropriate. For objects whose shapes can be approximated by rectangle or ellipse,
primitive geometric shape representations are more appropriate. For tracking objects with
complex shapes, for example, humans, a contour or silhouette based representation is
appropriate.

Fig 2, object representations (a) centroid, (b) multiple points, (c) rectangular patch, (d)
elliptical patch, (e) part-based multiple patches, (f) object skeleton, (g) complete object
contour, (h) control points on object contour, (i) object silhouette.
 CHAPATER 4
FEATURE SELECTION FOR TRACKING

Selecting the right features plays a critical role in tracking. The most desirable property of
visual feature is its uniqueness so that the objects can be easily distinguished in the
feature space. In general many tracking algorithms use these features. The details of
visual features are:

1. Color: The apparent color of an object is influenced primarily by two physical factors,
1) the spectral power distribution of the illuminant and 2) the surface reflectance
properties of the objects. In image processing, the RGB (red, green, blue) color space is
usually used to represent color.

2. Edges: Object boundaries usually generate strong changes in image intensities. Edge
detection is used to identify these changes. Algorithms that track the boundary of the
objects usually use edge as the representative feature.

3. Optical Flow: Is a dense field of displacement vectors which defines the translation of
each pixel in a region. It is computed using the brightness constraints, which assumes
brightness constancy of corresponding pixels in the consecutive frames.

4. Texture: Texture is the measure of the intensity variation of the surface whitch
quantifies properties such as smoothness and regularity.
 ALGORITHM FOR OBJECT TRACKING

Background subtraction in video using Bayesian learning:

An accurate and fast background subtraction technique for object tracking in still
camera videos. Regions of motion in a frame are first estimated by comparing the current
frame to a previous one. A sampling re-sampling based Bayesian learning technique is
then used on the estimated regions to perform background subtraction and accurately
determine the exact pixels which correspond to moving objects. An obvious advantage in
terms of processing time is gained as the Bayesian learning steps are performed only on
the estimated motion regions, which typically constitute only a small fraction of the
frame. The technique has been used on a variety of indoor and outdoor sequences, to
track both slow and fast moving objects, under different lighting conditions and varying
object-background contrast.
This algorithm presents robust system that achieves both (1) high speed and (2) high
degrees of sensitivity compared to existing techniques. To achieve these objectives a 2
step tracking system has been used.
1) Motion Region Estimation
2) Bayesian Sampling Resampling

Motion Region Estimation: The Block Matching Algorithm (BMA) is a standard way of
encoding video frames. A simplified variation of the BMA algorithm is used for
determining regions of each frame which have had motion relative to a reference frame.
Such regions have been called regions of motion. Each incoming frame is divided into
non-overlapping blocks of equal size. Each block is compared to the corresponding block
in the reference frame and the Sum of Absolute Difference (SAD) is determined for the
block. The reference frame may be chosen to be a few frames before the current frame, to
account for slow moving objects.
 Fig 1. Result of motion region estimation

Bayesian Sampling Resampling: A ‘sampling-resampling’ technique given by Smith

and Gelfand. This suggests easy implementation strategies and computational efficiency
while implementing Bayesian learning. Pixel observations at a particular spatial pixel
location are expected to form certain number of clusters. The parameters of these clusters
are thought to have probabilistic distributions of their own. These distributions are
updated via a Bayesian ‘Sampling-Resampling’ learning technique to obtain posterior
distributions..

Figure 2: The first row shows original frames from a video sequence. The second row
shows the results of motion region estimation. The third row shows the final Bayesian
Sampling-Resampling results.

Taxonomy of moving object detection algorithms:

 MOVING OBJECT DETECTION ALGORITHMS

1. FRAME DIFFERENCE:
In this method a background image without any moving objects of interest is taken as
reference image. Pixel value for each co-ordinate (x, y) for each color channel of the
background image is subtracted from the corresponding pixel value of the input image. If
the resulting value is greater than a particular threshold value, then that is foreground
pixel otherwise background. This method is simple and easy to implement, but the results
are not accurate enough, because the changes taking place in the background brightness
cause misjudgment.

1.1 An Improved Moving Object Detection Algorithm Based On Frame

Difference and Edge Detection
A combined approach by Zhan Chaohui is an efficient algorithm in which moving areas
are detected by forming several small blocks of edge difference image. The edge
difference image is obtained by computing difference between two images. Canny edge
detecting algorithm is used to detect the edges of continuous frames. The smallest
rectangle containing the moving object can be obtained. It is possible to get the exact
position of the moving objects by calculating connected components in binary images,
delete those components whose area are so small. The improved moving object detection
algorithm based on frame difference and edge detection has much greater recognition rate
and higher detection speed than several classical algorithms.
 Frame ( k-1) taken at time t frame (k) taken at time t+1

Edge map (EDGE k-1) of frame (k-1) edge map (EDGE k) of frame (k)

Edge difference image D(x, y) highlighting the difference of Edge maps of frame (k-1)
and frame(k).

Obtained motion areas are then mapped to the original image and appropriate edge pixels
(white pixels) are highlighted.
1.2 A Moving object Detection Algorithm for Smart Cameras

Yongseok Yoo suggested a new frame differencing method for moving object detection
using signed difference and Earth Mover’s Distance (EMD). First, a signed difference
image is acquired by subtracting two consecutive frames. For each fixed block in the
signed difference image, a motion pattern is calculated by EMD. The EMD is defined as
the minimum total amount of cost to move piles of earth to holes until all the earth is
moved or all the holes are filled. The neighboring blocks are then linked to detect moving
object regions.

The main idea behind this algorithm is to calculate matching costs for given
directions separately rather then to calculate exact EMD by linear programming. Here
block-based motion is used to locate moving object regions. An input image is divided
into blocks of fixed size and pairing vectors are calculated for each block. Blocks with
large pairing vectors indicate that there are motions in them. By combining these blocks,
moving objects can be detected.

Flow diagram:
 Original image Signed difference

Calculate motion strength Detect and Link motion blocks

Result
1.3 An Automatic Moving Object Detection Algorithm for Video
Surveillance Applications

Xiaoshi Zheng proposed an automatic moving object detection algorithm based on frame
difference and region combination for video surveillance applications. Initially an
automatic threshold calculation method is used to 0obtain moving pixels of video frames.
Frame difference is obtained by absolute difference value of two frames. Moving pixels
and static background pixels can be distinguished by a threshold value. In order to make
all moving pixels continuous and filter isolated pixels, moving regions are obtained by
morphological CLOSE operation.

In this algorithm we have three phases i.e. 1) moving object detection phase 2)
moving object extraction phase 3) moving object recognition phase

Moving object Detection Phase:

In order to detect a movement within a secured area, a surveillance camera is

positioned to monitor the area. The detection of a moving object within the monitored
area is the first phase. The movement detection uses a simple but efficient method of
comparing image pixel values in subsequent still frames captured every two seconds from
the surveillance camera. Two still images are required to detect any movement. The
Captured image size is 256x256 pixels. The first image is called “reference” image,
represents the reference pixel values for comparison purpose, and the second image,
which is called the “input” image, contains the moving object.
The two images are compared and the differences in pixel values are determined. If
the input image pixel values are not equal to the reference image pixel values, then the
input image. Pixel values are thresholded and saved in a third image, which is called
output image, with a black or white background. If the “difference” average pixel value is
smaller than a certain threshold value, then the output image background will be white
(pixel value is 255); otherwise, the background will be black (pixel value is 0).
After tracking the moving object motion, the previous input image will now be used
as a reference image, and a third image is captured and is called now the input image.
This process is repeated with the images being captured every two seconds, where the
same comparison method is applied. If there is a difference between the reference and
input images, then an output image is created. The obtained output image contains an
object that will be extracted (in the second phase).

Moving Object Extraction Phase:

The second phase is the extraction of a detected object. In this phase the output
image which was obtained at the end of the first phase is used. A simple but efficient way
of extracting the object is to use horizontal and vertical scanning of the output image. In a
two-dimensional matrix an object could be addressed by finding its vertical and
horizontal coordinates. The width and height of the extracted object image can be found
by using its starting coordinates and ending coordinates.
The implementation of this extraction method has two steps. Firstly, the output
image matrix is scanned horizontally, starting at the first x-direction coordinate value, the
pixel values in the corresponding column are summed up. The x direction coordinate
value is incremented by one and total pixel value in the next column is calculated. This
process is repeated until the last value in the x-direction. As result the total pixel values of
each column is calculated. Each total value is compared to a certain threshold value in
order to determine the x-coordinate where an object starts or ends within the image.
Secondly, the horizontal scanning method is repeated vertically, thus calculating the total
pixel value in each row, and then apply thresholding to determine the y coordinate where
an object starts or ends within the image. The background of the image that contains the
object is uniform as it has already been set to white or black at the end of the first phase.

Moving Object Recognition Phase:

This is the third and final phase. Here the detected and extracted moving object is
recognized by a trained supervised neural network that is based on the back propagation
learning algorithm. This algorithm is chosen due to its implementation simplicity and
efficiency in pattern classification.

2. BACKGROUND SUBTRACTION:
In this method, the moving regions are detected by subtracting the current image
pixel-by-pixel from a reference background image. The pixels where the difference is
above a threshold are classified as foreground otherwise background. Some
morphological post processing operations are performed to reduce noise and enhance the
detected region.

2.1 Real-Time Moving Object Detection for Video Monitoring System

The method of moving object detection is based on background subtraction for real time
moving objects. Guanglun Li proposes a new self-adaptive background approximating
and updating algorithm for moving object detection. To obtain the correct shapes of the
moving objects in every frame of the sequence, there are several steps. The subtraction of
two consecutive frames provides the image and the background model provides the
image. By using a temporal low-pass filter the background model is updated. The
updating process is applied to all the pixels of the model. In order to cope the sudden
changes with sudden light changes and leaves swings situation, finally AND/OR
operators are applied to the images to remove tiny noise in the images. The moving
object regions can extract accurately and completely by the self-adaptive threshold
segmentation method.

FLOW DIAGRAM:
 Fig 1 Fig 2

Fig 1 shows the shadow on right-top corner of the image.

Fig 2 shows the background image by using self-adaptive updating of background image.

EXPERIMENT:

Background image of 35th frame Background image of 105th frame

Background image of 175th frame

 Moving vehicle of 35th frame Moving vehicle of 105th frame

Moving vehicle of 175th frame

3. FRAME DIFFERENCE AND BACKGROUND SUBTRACTION:

The combination of background subtraction and frame differencing can improve the
detection speed and overcome the lack of sensitivity of light changes.

3.1 Moving Object Detection Algorithm Based On Improved

Background Subtraction

Lianqiang Niu describes an algorithm for moving object dtection based on combining of
improved background subtraction. This method can improve the detection speed and
overcome the lack of sensitivity to light changes. Considering the pixels relativity,
Gaussian Mixture Model in background subtraction is used. To extract a motion region,
the differences between the current frame and its previous frame is calculated.

After getting the motion scene background by improved Gaussian Mixture Model,
the foreground image is extracted. Foreground image is obtained by subtracting the
current image frame from background image. Symmetrical differencing is used to detect
the undetected regions. At each position of the pixel, the foreground images which are
achieved by using background subtraction and symmetrical differencing are processed by
a logical OR operation to obtain an accurate foreground image.

Original image results obtained after GMM

Foreground image masked original image

4. BACKGROUND UPDATING:

In background updating, the background of the selected pixels are replaced by the
average of the current and background pixels.

4.1 A Moving Object Detection Algorithm Based On Color Information

X H Fang suggested an algorithm to detect moving object based on color information.

This algorithm uses a pixel and its neighbors as an image vector to represent that pixel
and modeled different chrominance component pixel as a mixture of Gaussians, and set
up different mixture model of Gauss for different YUV chrominance components. In
order to make full use of the spatial information, color segmentation and background
model were combined. Simulation results show that the algorithm can detect intact
moving objects even when the foreground has low contrast with background.
In the spatial object surveillance systems, the detection of moving objects must be
quick and accurate. The background changes slowly in surveillance, so only detected
objects are usually considered to be moving. Hence the background model algorithm is
always used to detect moving object.
The principle of background model algorithm is to set up statistical model of
background, and then make the difference image of current image and background image
to extract moving foreground. Stauffer at al took use of Mixture of Gaussian (MOG) as
the statistic model of background [3], and every parameters of Gaussian distribution
change continuously to be adapt for the gradual change of background. The algorithm has
better adaptive capability for incomplete dynamic background. The fault of MOG is that,
when foreground texture and color are homogeneous and have low contrast with
background, the detected foreground is also not intact.
Shows the origin image, and the color of pedestrian’s pants is adjacent to the floor color.

Shows the result of origin MOG detection and legs which are overlap with floor can’t be
detected completely.

Shows the detected result of improved MOG, and legs which were overlap with floor can
be detected complete
 Shows the color image segmentation result with the edged image

Background model

Show the final detected result of joint color image segmentation and background
5. CROSS CORRELATION:

Manoj S Nagmode described a method to detect and track the moving objects to detect
and track the moving objects by using Normalized Cross Correlation algorithm (NCC). In
this an approach is proposed for the detection and tracking of moving object in an image
sequence. Two consecutive frames from image sequence are partitioned into four
quadrants and then the Normalized Cross Correlation (NCC) is applied to each sub frame.
The sub frame which has minimum value of NCC indicates the presence of moving
object. Next step is to identify the location of the moving object. Location of the moving
object is obtained by performing component connected analysis and morphological
processing. After that the centroid calculation is used to track the moving object. Number
of experiments performed using indoor and outdoor image sequences. The results are
compared with Simple Difference (SD) method. The proposed algorithm gives better
performance in terms of Detection Rate (DR) and processing time per frame.

5.1 A Novel Approach to Detect and Track Moving Object Using

Partitioning and Normalized Cross Correlation
Normalized cross correlation (NCC) algorithm is based on finding the cross correlation
between two consecutive frames in an image sequence. Correlation is basically used to
find the similarity between two frames. If the two consecutive frames are exactly same,
then the value of Normalized cross correlation is maximum. In that case no moving
object is detected. Now suppose there is a moving object in the image sequence, means
the two consecutive frames are not exactly same, with respect to positions of the pixel
values. In that case the value of Normalized cross correlation is less than maximum value
obtained. This concept of Normalized cross correlation is used for the detection of
moving object in an image sequence.
Flow diagram:
Experiments:

The tracking sequence of a walking person. This walking person is pointed by a red star.

Tracking sequence of multiple objects by simple difference method

Tracking sequence of multiple objects by PNCC method
Advantages and Disadvantages of Moving Object Detection Algorithms

Algorithms Advantages Disadvantages

Algorithm based on FD & Higher recognition rate and False detection under
edge detection higher detection speed complicated background

Algorithm for smart cameras Reject false motions due to Falsely detect specular
illumination changes reflections from moving
objects

Algorithm for video Automatic and efficient in Calculation is more complex

surveillance applications detecting moving objects

Real time moving object Extract moving object regions Processing time is strictly
detection for video accurately and completely depends on the quality of
monitoring systems moving points and on the
image dimension

Algorithm based on improved Increased running efficiency Object detected algorithm is

background subtraction and high detection accuracy complex

Algorithm based on color Detect foreground complexity Need improvements in real

information when foreground texture & time capability
 color are homogeneous

Partitioning & normalized Poor lighting conditions Average processing time per
cross correlation algorithm giving better results frame is high

CHAPATER

APPLICATIONS

Traffic information
SURVEILLANCE
 CHAPATER

CONCLUSION

Object tracking means tracing the progress of objects as they move about in visual scene.
Object tracking, thus, involves processing spatial as well as temporal changes. Certain
features of those objects have to be selected for tracking. These features need to be
matched over different frames. Significant progress has been made in object tracking.
Taxonomy of moving object detection is been proposed. Performance of various object
detection is also compared. It is not possible to consider a single method for all type of
images, nor can all methods perform well for particular types of image. The background
subtraction method detects object with noise and output is not accurate. Object behind
object is not detected. Problem occurs during identification of object when any obstacles
come before the object. If the position of camera is not proper and object in image is not
captured properly then it cannot be identified. To solve above problems and bring some
accuracy and richness by combining multiple methods and make use of it together
according to the application.

REFERENCES
[1] Rupali S.Rakibe, Bharati D.Patil,”Background Subtraction Algorithm Based Human
Motion Detection - published at: "International Journal of Scientific and Research
Publications (IJSRP), Vol. 3, Issue 5, May 2013 Edition".

[2] Himani S. Parekh, Darshak G. Thakore, Udesang K. Jaliya,"A Survey on Object

Detection and Tracking Methods", International Journal of Innovative Research in
Computerand Communication Engineering, Vol. 2, Issue 2, February 2014.

[3] Mr. Deepjoy Das and Dr. Sarat Saharia,” Implementation and Performance
Evaluation of Background Subtraction Algorithms”, International Journal on
Computational Sciences & Applications (IJCSA) Vol.4, No.2, April 2014.

[4] Thomas Andzi-Quainoo Tawiah,” Video Content Analysis For Automated Detection
And Tracking Of Humans In cctv Surveillance Applications”, School of Engineering and
Design Brunel University, August 2010.

[5] Qiang Zhu, Shai Avidan, Mei-Chen Yeh, Kwang-Ting Cheng,” Fast Human Detection
Using a Cascade of Histograms of Oriented Gradients”, June 2006.

[6] Navneet Dalal, Bill Triggs, and Cordelia Schmid,” Human Detection Using Oriented
Histograms of Flow and Appearance”, April 2006.

[7] Seema Kumari, Manpreet Kaur, Birmohan Singh,” Detection And Tracking Of
Moving Object In Visual Surveillance System”, International Journal of Advanced
Research in Electrical Electronics and Instrumentation Engineering (IJAREEIE) Vol. 2,
Issue 8, August 2013.

[8] Khushboo Khurana, Reetu Awasthi, “Techniques for Object Recognition in Images
and Multi-Object Detection”, International Journal of Advanced Research in Computer
Engineering & Technology (IJARCET) Volume 2, Issue 4, April 2013.

[9] Jae-Yeong Lee; Wonpil Yu, "Visual tracking by partition-based histogram

backprojection and maximum support criteria," Robotics and Biomimetics (ROBIO),
2011 IEEE International Conference on, vol., no., pp.2860, 2865, 7-11 Dec. 2011.
[10] L. D. Bourdev, S. Maji, T. Brox, and J. Malik,”Detecting People Using Mutually
Consistent Poselet Activations”, in Proc. EuropeanConf. on Computer Vision (ECCV),
pp. 168–181, 2010.