2005 IEEE International

Symposium on Signal Processing

and Information Technology

Motion Detection Using Differential Histogram Equalization

R. Hasanzadeh P.R.*, A. Shahmirzaie**, A.H. Rezaie***
Electrical Engineering Department,
Amirkabir University of Technology,
Hafez Avenue, Tehran, 15914, Iran.
Emails: [email protected]*; [email protected]**;[email protected]***

Abstract-In attention to general motion detection methods, ambiguous region which is between motion and silence
differential sequences of frames make some convenience regions to be realized as the class of movement (Fig. 1b).
which expand it’s usability for lots of applications. In Comparing Fig. 1a with Fig. 1b, we notice that the Bayes test
practice, differential sequences of frames have been affected for minimum cost is a likelihood ratio test with a different
by environmental noises and CCD cameras structure that threshold from Bayes test for minimum error [11].
make difficulty detection in volubility of objects. In this In these systems the second priority after the motion
method using differential histogram equalization, we try to detection is the speed of data processing. Increasing the speed
enhance and increase the contrast between data and noise of process causes an immediate result to the expert operator
histogram levels. The new threshold levels adaptively and the central control system. Therefore although accurate
obtained by simple algorithm and doesn’t need to know motion detection may be achieved by iterative methods,
probability distribution function of the noise. because of their long calculations time may not be used in
real time systems.
Keywords– Motion Detection, Histogram Equalization, Adaptive
Threshold. II. METHOD DESCRIPTION

I. INTRODUCTION Detection of motion in images can be obtained by

investigating of frames. One of the methods for motion
Motion detection based on video signals is widely used in detection is based on using of two sequential frames.
security systems, airports and military bases etc. [1, 2 and 3]. Differential histograms of two frames contain Gaussian noise
The situation is usually the detection of the movement of an with zero mean caused of CCD camera sensors [9]. In
object in a constant background. In other words it is assumed practice, its variance has small variation specifically for
that the position of the camera is fixed. Several methods are objects that contain motions smaller than 1/8 of space area.
used to motion detection, such as Dynamic Time Warping, Fig. 2 shows a differential histogram without motion which
Hidden Markov models and Neural Networks [12], but usual obtained in practice. Here we use a gray scale CCD camera
method is to use the two consecutive frames having the
with 256 brightness levels (8 bit).
knowledge of the probability density function of the noise,
possible to extract some sections of the differential histogram
levels which have the most probability of the helpful data [3,
and 4].

Fig. 1a. Fig. 1b.

In these methods the necessity of having the knowledge of Fig. 2. Shows Differential histogram without motion
environmental noise, limits its wide uses in automatic control
As a result of discrete brightness level problem and small
systems [5, and 6]. Efforts of applying any adaptive method
variation of variances, the number of fundamentals levels of
for the threshold levels involve algorithms like K- Means
which are time consuming [8]. Usually high precision differential histograms which contains at least 5% of
boundary detection using likelihood ratio is not essential in amplitude of the level at zero mean range is less than twenty.
security systems (Fig. 1a) but also pattern recognition in Hence, obtaining the variance of Gaussian distribution
security systems is based on motion detection assumed on function has some problem. The number of threshold level
motion existence. On the other hand it is preferred the (handled or adaptive) used for extract motion from noise, is
more than 10. In differential histogram space nearly all of

0-7803-9314-7/05/$20.00©2005 IEEE 186

variations are between [-100,100]. The motions comparable might neglect remarkable volume of data or impose noise to
with whole of image space, for example ¼ or ½ of entire area motion detection. Fig. 5 shows DHE before and after motion
of image are detected without any complex algorithms only for Fig. 4. Probability of unwanted oscillation’s effect is
use a simple threshold. Generally, the complexity of minimized by high contrast of brightness level. Aggregations
problems is obtaining a solution to detection of nearly small of brightness levels in two corners of DHE belong to motion
motions in image space because differential histogram levels and in centre of DHE with high distance between levels
of motion concentrated around of differential histogram belong to noise.
levels of noise.

Fig. 5a. DHE before motion

Fig. 3. Shows Differential histogram with motion

Hence, specification of threshold levels is very sensitive
which is shown in Fig. 3. The Fig. 3 shows a differential
histogram (DH) levels which can be extracted in practice.
The TR and TL, point to the threshold levels that one can set
handled or adaptive to separate the motion from the noise and
Fig. 5b. DHE after motion
the background.

III. DIFFERENTIAL HISTOGRAM EQUALIZATION

Fig. 4a. DH before motion

Fig. 6. Shows stages of

Fig. 4b. DH after motion motion detection
Also in practice the DH have some oscillations around the Attention to Fig. 6, video capture system provides images
zero mean level that is due to the fast motions. Differential then these images are used in sequences of differential frames
histogram equalization or DHE increase image enhancement obtaining differential histograms. Histogram equalization
and contrast of brightness levels and remove threshold level [10] uses for enhancing brightness levels of one image but
variation is caused by DH oscillation (See equation (3)). The here we applies it for enhancing differential brightness levels
Fig. 4 shows a DH before and after motion in image. of sequential frames.
Differential levels continuously have a nonzero value in all of Percentage of each brightness level computed by equation
(1):
the levels between [-20, 20] and after motion between [-
100,100]. If DH threshold levels changes only one level, it n
Pr [K] = k −255 ≤ K ≤ 255 (1)
n

187
n k : The number of pixels with K differential brightness sequential images which were obtained from video capture
system. The simple but reliable and adaptive thresholding by
level. using DHE can allow us to real-time surveillance. Fig. 8a
k nj shows the original image which is obtained from video
Sk = CDF[k] = ∑ (2)
j= 0 n capture system. The speed of video capturing system was 25
Sk , cumulative distribution function, is new differential frames per second that in addition to processing decreases to
20 frames per second. The result of difference of two
brightness level. Percentage of each new brightness level sequential images is shown in Fig. 8b. In order to find the
computed by equation (3): region of the moving object it is necessary to determine a
closed contour around the detected motion. This may be done
Ps [K.Sk ] = ∑ Pr [K.S j ] by first the differentiation of two consecutive frames
j (3) following DHE calculations which is shown in Fig. 8c and
for [K.S j ] = [K.Sk ] applying the algorithm mentioned in section (IV) which
results to Fig. 8d. The number of nonzero continuous
Equation (3) converts a number of the adjacent brightness brightness levels (ZL and ZR) that is used here are two. The
levels that occurred with small amplitude, to one brightness some simple and small noise which is remained can be
level with remarkable amplitude. Different between Figs. 4 removed by morphological processing [10]. It is done by two
and 5 shows it. stages of Morphology algorithm (Dilation/Erosion) followed
by using the remove particle algorithms will lead to figure (8-
e). In this case a uniform and binary region of the moving
IV. ADAPTIVE THRESHOLD object will be achieved by using the filling algorithm as
shown if Fig. 8f.
As a previous discussion, motions contain big area and Also it is necessary to find out a rectangle around the
motions with high contrast relative to background doesn’t moving object. The number of rectangles depends on the
need to complex or high precision algorithm because these number of motion detected regions in each camera which is
motion can be detected simply. The problem is detection of shown in Fig. 8f.
motions with brightness levels near to background or motions It is not necessary to transmit the entire image, because the
with small area. Two last properties cause to: (1) low background of the images is constant. We just extract the
moving images from the boundary that was obtained as Fig.
amplitude of motion brightness levels, (2) ling motion
8f. The results of the moving objects from the Fig. 8a are
brightness level near the center of DHE which the noise
shown in Figs. 8g and 8h. Finally, extracted images from
levels stand here. After DHE which is shown in Fig. 5, we try
moving object are sent through LAN and/or Modem to
to detect a threshold adaptively. The thresholds calculated as Central Security System.
nonzero continuous brightness levels with a number of
specific zero amplitude level after them. Because of small VI. CONCLUSION
and low variation, after DHE brightness levels of motion
have a continuously state at the corner of histogram. In some application such as security systems it is necessary
Threshold between motions levels and noise can be using of methods which are fast, reliable and simple. Usually
calculated by counting the continuous brightness levels with high precision boundary detection using likelihood ratio is
zero amplitude between them. Numbers of allowable zero not essential in security systems but also detection of motion
amplitude levels depend to motion area and brightness is vital. In this method using differential histogram
variation. Usually, it is equals two. equalization, we try to minimize the cost for the motion class
by enhancing and increasing the contrast between motion and
Fig. 7 shows the flowchart for obtaining the thresholds. H iR
no-motion histogram levels. The new threshold levels
Shows the Right side of DHE and H Lj shows the Left side of adaptively obtained by simple algorithm and doesn’t need to
it { i ∈ (0, 255], j ∈ [−255, 0) }. TR and TL are Right and Left know probability distribution function of the noise. This
threshold levels of DHE. ZR and ZL shows continuous method is applied for monitoring and security system of
brightness levels with zero amplitude from the two corners Electrical and Mechanical Eng. Departments of Amirkabir
re-entrant to centre of DHE. X is the numbers of allowable University (10 floors and 30 cameras). The result of it had
continuous levels with zero amplitude that usually is 2. good efficiency for fast detection of motion and cause to
reduce the human inspection fault.
V. EXPERIMENTAL RESULT

We apply the method for detection the motion in PC based

security systems. The size of images was 320*240 in eight bit
Gray scale format. The motion has been detected from two

188
 ACKNOWLEDGEMENT

The authors would like to thank Mr. Javad Ostadreza,

Administrator of Computer Networks, Department of
Electrical Engineering, for helping to design the monitoring
and security system of Electrical and Mechanical Eng.
Departments, Amirkabir University of Technology.

REFERENCES
[1] McKenna, SJ. , Jabri, S., Duric, Z., Rosenfeld, Wechsler, H., “Tracking
groups of people” Int. J. Comput. Vision Image Understanding, vol. 80, pp.
42-56, 2000.
[2] Sukthankar, R., Stockton, R., “Argues: the digital doorman” IEEE Trans.
Intell. System, vol. 16(2), pp. 14-19, 2001.
[3] Kim, J.B., Lee, C.W., Hwang, S.W., Kim, H.J. “A real time moving
object detection for video monitoring system” Proc. ITC-CSCC 1, pp. 454-
457,2001b.
[4] Nariman, H., Alireza, M., Neil, B., “Automatic thresholding for change
detection in digital video” Proc. SPIE 4067, pp. 133-142, 2000.
[5]Kittler, J. Illingworth, J., “Minimum error thresholding pattern
recognition”, vol. 19 (1), pp. 41-47, 1986.
[6] Aach, T., Kaup, A., Mester, R., “Statistical model-based change
detection in moving video” Signal Processing, Vol. 31, pp.165-180, 1993.
[7] Kim, M., Chai, J.G., Kim, D., Lee, H., Lee, M.H., Ahn, C., Ho, Y.s., “A
vop generation tool: Automatic segmentation of moving objects in image
sequences based on spatio-temporal information” IEEE Trans. Circuits Sys.
Video Tech., Vol. 9, no. 8, pp. 1216-1226, 1999.
[8] Kim, J.B., Kim, H.J.,”Efficient region-based motion segmentation for
video monitoring system” Pattern recognition, vol. 24 (3), pp. 113-128, 2003.
[9] Teklap, A.M., Digital video processing, Prentice Hall, upper saddle river,
NJ, 1995.
[10] Gonzalez, R. C., and Woods, R. E., Digital Image Processing, Addison-
Wesley Publishing Company, Massachusetts, 1992.
[11] K. Fukunaga, Introduction to Statistical Pattern Recognition, Academic
Press, INC, 1992. Fig. 7. Shows the flowchart to
[12] L. Wang, W. Hu, T. Tan, “Recent Development in human motion achieving to thresholds
analysis”, Pattern Recognition, vol. 36, pp. 585-601, 2003.

189
 Fig. 8a. The Original Image Fig. 8b. The result of difference of two Fig. 8c. After Differential Histogram
sequential images Equalization

Fig. 8d. The result of adaptive thresholding Fig. 8e. Particle removing Fig. 8f. Motion boundary detection

Fig. 8g. First moving object Fig. 8h. Second moving object

190