You are on page 1of 4

Object Detection with improved Object Localization

for HOG and SVM


1st Kalyan S 2nd Ajit Kulkarni
dept. of Electrical Engineering Senior Technical Architect
National Institute of Technology,Calicut Continental Automotive
India Banglore, India
kalyan m200077ee@nitc.ac.in ajit.2.kulkarni@continental-corporation.com

Abstract—Object localization is a key operation in computer (or boundaries) of the image.It is one of the basic steps in
vision tasks such as Object detection, where applications depend image processing, pattern recognition in images and computer
on the rapid and accurate position estimation of objects.An vision. When we process very high-resolution digital images,
Improper Bounding Box coordinates results in inaccurate dis-
tance calculation of the object from the camera.There are CNN convolution techniques come to our rescue.
based localization methods such as pose estimation algorithm or
grabcut technique but they require more memory.We present a There are various methods in edge detection, and the
simple and efficient edge detection based localization algorithm following are some of the most commonly used methods [2]
that corrects the bounding box position.This involves determining -Prewitt edge detection,Sobel edge detection,Laplacian edge
all the edge lines of the object and recalculating the bounding detection,Canny edge detection.We have used prewitt edge
box coordinates.
Index Terms—Bounding Box, Localization, Object detection, detector [3] as it is the simplest of all the edge detecting
Hough Transform methods.
V prewitt kernel =
I. I NTRODUCTION  
-1 -1 -1
Object detection is a multi-task learning problem with the 0 0 0
goal of correctly identifying the object in the image while 1 1 1
also localizing the object into a bounding box, therefore the
end result of the object detection is to classify and localize
the object.After determining the bounding box coordinates
[tlx, tly, brx, bry],the next step is to determine the distance B. Hough Transform
of the object from the camera. Hough transform is a feature extraction method used in im-
age analysis. Hough transform can be used to isolate features
The distance is calculated based on the foot of the object
of any regular curve like lines, circles, ellipses, etc. Hough
i.e.,((trx + brx)/2, bry), the midpoint of the lower horizontal
transform in its simplest from can be used to detect straight
line of the bounding box.Therefore it is highly important to
lines in an image. For our analysis we use line detection with
calculate the distance properly,if not there will be an accident.
Hough Transform.The Hough Line Transform is a transform
Some times only a certain portion of the entire object used to detect straight lines. To apply the Transform, first an
may fit into the bounding box or the bounding box may be edge detection pre-processing is desirable.
much bigger than the object inside(loose bounding box).We
need to make sure that the object fits tightly in the bounding In an image analysis context, the coordinates of the point(s)
box.So,We propose a method to get a tighter bounding box of edge segments (i.e. X, Y ) in the image are known and
around the object by finding the edges of the object and using therefore serve as constants in the parametric line equation,
Hough Transform [1] to determine the exact coordinates of while R(ρ) and Theta(θ) are the unknown variables we seek.
those edge lines.This Paper mainly focuses on the pedestrian If we plot the possible (r) values defined by each (theta), points
localization. . in Cartesian image space map to curves (i.e. sinusoidal) in the
II. R ELATED W ORK polar Hough parameter space. This point-to-curve transforma-
tion is the Hough transformation for straight lines.
A. Edge Detection
Edge detection is a technique of image processing used to
identify points in a digital image with discontinuities, simply
III. A LGORITHM
to say, sharp changes in the image brightness. These points
where the image brightness varies sharply are called the edges Below is the Algorithm for Bounding Box correction
Fig. 5. After thresholding
Fig. 1. Before the Bounding Box Correction

Fig. 6. After Hough Transform

Fig. 2. Old Bounding Box

Fig. 7. After the Bounding Box Correction


Fig. 3. After Blurring

Fig. 4. after Prewitt filter Fig. 8. Before Bounding Box Correction(cropped pedestrian)
Algorithm 1 Bounding Box Correction Algorithm
Require: a tighter Bounding Box
Ensure: bry ≈ y2(bottom most line)
set flag to 1
while flag 1 do
1. get bounding box(tlx,tly,brx,bry) image
2. apply Gaussian blur
3. apply prewitt filter
4. apply binary threshold
5. apply hough transform
if (bry > bottom most line’s y2) then
Fig. 12. After thresholding)
1. bry = y2(bottom most line)
2. flag = 0 bry = y2 (bottom most line)
1. bry = bry + 5
end if
end while

Fig. 13. New bounding Box)

A. Flow of Algorithm
B. Explanation
Fig. 9. Old Bounding Box)
The Bounding Box Correction is a necessary step to get
correct distances of objects from the camera.The first step is to
get the current bounding box image data as shown in fig 2.We
need to perform correction operation for a bigger bounding
box as well as the one where the whole object is not inside
the bounding box.The second step is to apply Gaussian filter
to smooth out the image to reduce noise as shown in fig 3.The
third step is to apply prewitt filter.

We have chosen the prewitt filter because ,it is the simplest


filter of all the filters to reduce complexity.The fourth step is

Fig. 10. After Smoothing)

Fig. 11. After Prewitt filter)


Fig. 14. After Bounding Box Correction)
to perform Binary Threshold so that only edges are visible
as shown in fig 4.Now we don’t have the coordinates of the
edge lines.This is where we use Hough Transform to get the
coordinates of each of the edge lines.The fifth step is to see
whether the edges are far from the bottom line of the bounding
box(bigger box case) or the edges touch the bottom line of the
bounding box(cropped box case) as shown in fig 5.

If the box is bigger ,bry is made equal to y2 of the last


line which is near the bottom of the bounding box.If the box
is a cropped one,bry is increased by 5 pixels and the process
is repeated until the bry is very close to the last line’s y2 as
shown in fig 7

R EFERENCES
[1] R. O. Duda and P. E. Hart, “Use of the hough transformation to detect
lines and curves in pictures,” Commun. ACM, vol. 15, pp. 11–15, 1972.
[2] D. Ziou and S. Tabbone, “Edge detection techniques - an overview,”
vol. 8, 06 2000.
[3] G. Shrivakshan and C. Chandrasekar, “A comparison of various edge
detection techniques used in image processing,” International Journal of
Computer Science Issues, vol. 9, pp. 269–276, 09 2012.

You might also like