There good response with movement tracking [4-5]. Parts

There are different approach to estimating the people count from an image and videos.
Some of the approaches are density based; some of them are background subtraction based
and some of them are corner based people detection 2. Density based energy minimization
framework which combines crowd density estimates with the strength of individual person
detection 3 and tracking. Background subtraction based people detection which has a lower
accuracy but having a good response with movement tracking 4-5. Parts based object and
crowd detection and tracking in still images have some research 6. Scene priming to constrain
locations of objects in the image 7. Constant scale model, sliding window pyramid and scale
window word density based people and parts of body detection as well as head detection is
described in 6. This paper describes a pedestrian detection system that integrates image
intensity information with motion information. We use a detection style algorithm that scans a
detector over two consecutive frames of a video sequence. The detector is trained (using
AdaBoost) to take advantage of both motion and appearance information to detect a walkingperson. Past approaches have built detectors based on motion information or detectors based
on appearance information 8 also development of a representation of image motion which is
extremely efficient and implementation of a state of The art pedestrian detection system which
operates on low resolution images under difficult conditions have analyzed 8. Background
subtraction results based estimation the number of people in a complicated scene which
includes people who are moving only slightly an Expectation Maximization (EM)-based method
has been developed to locate individuals in a low resolution scene 2. A global data association
method based on Generalized Graphs for tracking each individual in the whole video. In videos
with high crowd-density, Track individuals Using a scene structured force model and crowd flow
modeling 1. Contextual information without the need to learn the structure of the scene based
people detection 1. Integral channel features for image classification tasks, focusing in
particular on pedestrian detection 9 that focuses on feature extraction and feature description
from RGB images. Integral channel features is that multiple registered image channels are
computed using linear and non-linear transformations of the input image, and then features
such as local sums, histograms, and Haar features and their various generalizations are
efficiently computed using integral images 9. Compute crowd density maps in order to estimate
the spatial distribution of people in the scene also the self adaptive dynamic parameterization
based people detection 10 basically that works based on human aspect ratio.
There are some approaches based on density, background subtraction that we
mentioned in chapter two. There are different approaches of feature extraction for human,
pedestrian or crowd detection. Some of the approach is integral channel feature, corner feature,
are measurement, energy function behavior. Our approach is orientation of gradients based
object area locating by defining the interest point, based on these points we have removed
some of the points that are overlapping or not import. For selecting or detecting the interest
points we have applied an RGB image to Gray image conversion. From gray image we have
calculated the orientation and gradients in X, Y direction. Based on that orientation value we
have defined a range maximum and minimum to get a binary image. From this binary image we
can decide separate object region by applying connected component algorithm 11. Before this
step we have marked some of the area that is not important for our computation, as example
the region that consisted with 20 pixels only we have discarded these regions. Removing this
regions we get another binary image, in this image for finding a separate region, applied some
connected component algorithm 11 here 8-connectivity is used. This step provides a binary
image with the region information.
Now it’s turn to get some concrete value from each of the region, we have selected
centroid value as concrete value of each region. Based on this centroid value we have selected
a patch area of 201×201 from original RGB image.
Next steps is feature extraction, there was some previous approach like density,
integrated channel feature and histogram etc. We have selected the Histogram Oriented
Gradient (HOG) as a feature extraction technique. HOG is very much responsive to head or
face detection because its computes the orientation changes based on color information
changes that described more on chapter four. We have trained the SVM by our manual
annotated dataset positive and negative both of the set at a time. Tested each of the HOG value
with SVMStruct that means classify that point with trained SVM. Get the result from it. This
result indicate the crowd face is detected. In case of non face and overlapping face this system
also showing marked area as crowd people. Hence gradient information based binary image
creation and the HOG feature based crowd face gives us a good result that’s around 90%
accurate detection.