DETECTING PERSONS USING HOUGH CIRCLE TRANSFORM

IN SURVEILLANCE VIDEO

Hong Liu, Yueliang Qian and Shouxun Lin

Key Laboratory of Intelligent Information Processing & Research Center of Pervasive Computing

Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China

Keywords: Person Detection, Hough Circle Transform, Elevator Surveillance.

Abstract: Robust person detection in real-world images is interesting and important for a variety of applications, such

as visual surveillance. We address the task of detecting persons in elevator surveillance scenes in this paper.

To get more passengers in the lift car, the camera usually installed at the corner of ceiling. However, the

high and space of lift car are limited, which makes person occluded by each other or some parts of body

invisible in captured images. In this paper, we propose a novel approach to detect head contours, which

includes three main steps: pre-processing, head contour detection and post-processing. Hough circle

transform is adopted in the second stage, which is robust to discontinuous boundaries in circle detection.

Proposed pre-processing and post-processing methods are efficient to remove false alarms on background or

body part. Experimental results show our proposed approach is time saving and has better person detection

results than some other methods.

1 INTRODUCTION

Robust person detection in real-world images is

interesting and important for a variety of

applications, such as visual surveillance and driver

assistance systems. For surveillance applications,

examples include the estimation of queue length in

retail outlets, the monitoring of entry points, bus or

elevator terminals. Persons are one of the most

challenging categories for object detection. A large

variability in their local and global appearance is

caused by various poses or styles of clothing (Leibe,

2005). Many techniques (Giebel, 2004, Schmid,

2004) have been made to detect in constraint

settings. Often it is assumed that the people in the

scene are well separated. While in many real

applications several persons may be present in the

same image region, partially occluding each other,

which bring difficulty to detecting and counting

them.

Various methods have been applied to person

detection for surveillance applications. Among them,

classification techniques by supervised learning can

be applied to decide if a given image region contains

a person. For example, (Nakajimia, 2000) uses

Support Vector Machines to recognize people.

(Gravrila, 2000) uses a tree based classifier to

represent possible shapes of pedestrians. (Viola,

2001) proposes a framework for face detection,

which use Haar wavelets and AdaBoost to build a

cascade structured detector. However all above

approaches requires certain training data to learn the

parameters of the underlying model. Some other

methods use contour or edge feature to detect

person. (Zhang, 1995) introduces a Model-specified

Directional Filter (MDF) to identify the pedestrian

on an escalator by extracting the contour feature of

pedestrian’s head in an image with the camera

installed right above. However, neither of above

papers addresses the situation of crowded scenes

where persons may be occluded by each other.

In this paper, we specifically address the task of

detecting persons in elevator surveillance scenes.

The goal is to detect all pedestrians, localize them

and count them in the image. With the number of

passengers, the elevator control system can

minimize waiting time and ride time of passengers

and reduce the unnecessary elevator stop, which can

finally save electronic energy and reduce

maintenance fee of elevator. For elevator

application, the high and space are limited in lift car.

The camera usually installed at the corner of ceiling

in lift car to capture more passengers, which will

deform the appearance of person and make some

267

Liu H., Qian Y. and Lin S. (2010).

DETECTING PERSONS USING HOUGH CIRCLE TRANSFORM IN SURVEILLANCE VIDEO.

In Proceedings of the International Conference on Computer Vision Theory and Applications, pages 267-270

DOI: 10.5220/0002856002670270

 SciTePress

Head contour

detection by

Hough circle

transform

pre-processing

post-processing

Remove head

contour by

color feature

Remove head

contour by

foreground

image

Edge of adaptive

background image

Edge of

current image

Edge of

foreground

Get the final

head contour

Figure 1: The framework of our approach.

parts of person invisible. On the other hand, persons

often occluded by each other and pose of head are

various in captured images.

Although detecting and counting persons in

elevator application is a very difficult task, we can

also see that the appearance of person head is similar

as circle. In this paper, we propose a framework to

detect person head based on Hough circle transform

(HCT), which is proved useful in occluded situation.

Our approach is given in section 2. Experimental

results of the detecting system are presented in

section 3. Conclusions are given in section 4.

2 OUR APPROACH

2.1 Outline of Our Approach

For person detection in elevator surveillance scene,

we propose a novel method as Figure 1 shows. The

approach consists of three main steps: pre-

processing, head contour detection and post-

processing.

HCT algorithm is used in the second stage,

which is based on the edge points in the images. So

in order to reduce some interference by edge points

on background field, certain pre-processing is

needed. In pre-processing step, Canny edge

detection is performed on adaptive background

image and current frame. Then subtraction is used to

get the edge of foreground, which can remove many

edge points on background.

After above process, HCT based on gradient

(Yuen, 1990, Bradski, 2008) is performed on edge

image of foreground to get quasi circles. This step is

detailed in section 2.2. After this step, some image

regions satisfied HCT are decided as circles, which

can be see as candidates of head contours. However,

with the interference from background and clothing,

some candidates are not related with person head.

Finally, post-processing is adopted to remove some

false alarms, which is detailed in section 2.3. After

above three stages the final result of head contours

can be obtained.

2.2 Hough Circle Transform

The Hough circle transform (HCT) shows

considerable immunity to problematic object

boundaries, such as ones partially obliterated by

occlusion, overlapping effects and breakages, as

well as boundaries distorted by interference noise,

ambient illumination, object motion (Yuen, 2001).

(Kimme, 1975) gives probably the first known

application of the Hough Transform to detecting

circles in real images. In their work, they have made

use of the direction of the gradient at each edge

point. If a circle in the image is described as

222

()()

aybr



 

(1)

where

(,)ab

are the coordinate of the circle

center and

r is its radius, then an arbitrary edge

point

(, )

will be transformed into a right

circular cone in the

(,,)abr

parameter space. If all

the image points lie on a circle then the cones will

intersect at a single point in

(,,)abr

corresponding

to the parameters of the circle. This would mean far

greater memory requirements and much slower

speed. (Yuen, 2001).

To solve these problems, Hough gradient method

is propose in (Yuen, 2001). First the image is passed

through an edge detection phase. Next, for every

nonzero point in the edge image, the local gradient is

considered. Using this gradient, every point along

the line indicated by this slop from a specified

minimum (

min_ r

) to a specified maximum

(

max_ r

) distance is incremented in the

accumulator.

For elevator application, color of person head

may by darker or lighter than that of ambient

background as figure 2 shows. So we remain two

directions of slope to get candidate centers in our

paper.

At the same time, the location of every one of

these nonzero pixels in the edge image is noted. The

candidate centers are then selected from those points

in this accumulator that are both above some given

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

268

threshold

_acc thr

and larger than all of their

immediate neighbours. These candidate centers are

sorted in descending order of their accumulator

values, so that the centers with the most supporting

pixels appear first. For each center, all of the

nonzero pixels are considered. These pixels are

sorted according to their distance from the center.

Working out from the smallest distances to the

maximum radius, a single radius is selected that is

best supported by the nonzero pixels according to

the threshold

_Rthr

. A center is kept if it has

sufficient support from the nonzero pixels in the

edge image and if it is a sufficient distance from and

previously selected center. (Bradski, 2008).

So we can see Hough gradient method can be

decomposed the circle finding problem into two

stages. First stage is to find candidate centers and

second stage is to find satisfied radius of circle. A

two dimensional array is required to accumulate

votes along the normal of each edge point. To

identify the radius of circles, the distance of each

point from a candidate centre is calculated and a

radius histogram is produced. The storage space

required for the method is quite small, since only a

single 2-D accumulator and a 1-D histogram are

necessary. (Yuen, 2001).

This Hough gradient method run much faster

and, more importantly, helps overcome the problem

of the otherwise sparse population of a three

dimensional accumulator. (Bradski, 2008).

2.3 Post-processing

After above step, some image regions satisfied HCT

are decided as head candidates. While with the

interference from background and clothing, some

candidates are not related with person head.

Although edge subtraction of background is

taken in pre-processing step, some circles at

background are also detected after HCT process. So

our post-processing should remove candidate circles

according to foreground image firstly.

For elevator application, pose of head are usually

various in images. Sometime we can see back, side

or front of head. So the color of head is usually

different according to the pose of head. While

without regard to wearing hat, the color distribution

of head can be clustered. In this paper, we use GMM

to get the color clusters around the center of head in

marked images. To reduce the illumination

interference, we transform the image from RGB

color space to HSV color space. Here, H (hue) and S

(saturation) values are used to color cluster.

In post-processing, we check the color feature of

the region beside candidate center to decide whether

this circle can be maintained. After this process,

some candidates with different color feature as head

can be removed.

3 EXPERIMENTAL RESULTS

To compare our approach, we also realize another

method based on Adaboost classifier (Viola, 2001).

The train data come from videos captured in lift car.

We use 783 images with 1788 head-shoulders as

positive samples and 8000 images without person as

negative samples collected from Corel Image

Gallery. All manually marker samples are

normalized to the size of 20×20 pixels. We train one

classifier with 16 stages of cascade. Some detect

results from this method are shown in figure2. The

red rectangles are the results after head-shoulder

detection based on Adaboost classifier. We can see

that some false detections on body and background,

while many head regions are missing in the second

image in crowed situation.

Figure 2: Some results based on Adaboost.

We also test our approach to detect head contour

in elevator surveillance video. Some results are

shown in figure 3. Figure 3(a) is the result only

using HCT method. The green points are the centers

of circle and the red circles are the corresponding

head candidates. Many head candidates are false

alarms in this image, some come from regions on

background and some located on body part, which

have different color feature compare with usual head

color. So we adopt pre-processing and post-

processing to modify this condition.

Figure 3(b) is the edge image of adaptive

background image after Canny edge detection. Edge

image of current image is also obtained. Then

subtraction is performed to get the edge of

foreground as figure 3(c) shows, which reduce some

interference edge points at background. Finally,

figure 3(d) shows the result using our method. With

pre-processing and post-processing, some false

detection on background and body are removed.

Figure 4 shows another result based on our approach

DETECTING PERSONS USING HOUGH CIRCLE TRANSFORM IN SURVEILLANCE VIDEO

269

and the corresponding values of some parameters are

also shown in annotation of image. From the

experimental results, we can see our approach has

better performance than some usually methods for

elevator surveillance application.



(a) Result from HCT (b) Edge of background



Figure 3: Some results based on our approach.

(

_acc thr

=25,

min_ r

=10,

max_ r

=40,

_Rthr

=60).



(a) Result from HCT (b) Result from our method

Figure 4: Some results based on our approach.

(

_acc thr

=25,

min_ r

=10,

max_ r

=35,

_Rthr

=50).

4 CONCLUSIONS

In this paper, we present an approach for person

detection in elevator surveillance application. To get

more passengers in the lift car, the camera usually

installed at the corner of ceiling. The high and space

of lift car are limited, which makes person occluded

by each other or some parts of body invisible in

captured images. So it is a difficult task to detect

person in this scene.

In this paper, we propose a novel method to

detect head contours in surveillance video. The

method includes three main steps: pre-processing,

head contour detection and post-processing. Hough

circle transform is adopted in head contour detection

stage, which is shown robust to discontinuous

boundaries in circle finding. Pre-processing can

reduce some edge points on background, which can

remove some interference. Hough gradient method

is efficient and time saving for circle detections in

the head contours detection step. The last step post-

processing can remove many false head contours on

background or on body part that has different color

distribution as usual color of head. The experimental

results show our approach has better detection

performance than many other methods. In the future,

we will test our approach on large dataset.

ACKNOWLEDGEMENTS

This work was supported by the National Nature

Science Foundation of China (60802067).

REFERENCES

Leibe, B., Seemann, E., Schiele, B., 2005, Pedestrian

Detection in Crowded Scenes. In Proc. Computer

Vision and Pattern Recognition, vol.1, 878-885.

Gavrila, D., 2000, Pedestrian detection from a moving

vehicle. In Proc. 6th European Conf. Computer Vision,

Dublin, Ireland, vol.2, 37–49.

Giebel, J., Gavrila, D.M., 2004, A Bayesian framework for

multi-cue 3d object tracking. In Proc. 8th European

Conf. Computer Vision, Prague, Czech Republic, 241–

252.

Schmid, C., Mikolajczyk, K. and Zisserman, A., 2004,

Human detection based on a probabilistic assembly of

robust part detectors. In Proc. 8th European Conf.

Computer Vision, Prague,Czech Republic, vol.1, 69–

82.

Nakajima, C., Pontil, M., Heisele, B.and Poggio, T., 2000,

People recognition in image sequences by supervised

learning. In MIT AI Memo.

Viola. P. and Jones. M., 2001, Rapid object detection

using a boosted cascade of simple features. In IEEE

Conference on Computer Vision and Pattern

Recognition.

Zhang, X.W., Sexton, G., 1995, A new method for

pedestrian counting. Proceedings of the Fifth

International Conference on Image Processing and its

Applications, 208- 212.

Yuen, H. K., Proncen, Illingworth, J., Kittler, J., 1990,

Comparative Study of Hough Transform Methods for

Circle Finding. Image and Vision Computing, Vol.8,

No.1, 71-77.

Kimme, C., Ballard, D. H. and Sklansky, J., 1975, Finding

circles by an array of accumulators. Communications

of the Association for Computing Machinery, vol.18,

120-122.

Bradski, G., Kaebler, A., 2008, Learing Opencv:

Computer vision with the Opencv library. Publisher:

O'Reilly Media.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

270