Reliable Moving Vehicle Detection Based on the Filtering of Swinging Tree Leaves and Raindrops

(1)

Reliable moving vehicle detection based on the ﬁltering of swinging tree leaves

and raindrops

Deng-Yuan Huang

a,⇑

_{, Chao-Ho Chen}

b

_{, Wu-Chih Hu}

c

_{, Sing-Syong Su}

b

a

Department of Electrical Engineering, Dayeh University, 168 University Rd., Dacun, Changhua 515, Taiwan, ROC b

Department of Electronic Engineering, National Kaohsiung University of Applied Sciences, 415, Chien Kung Rd., Kaohsiung 807, Taiwan, ROC c

Department of Computer Science and Information Engineering, National Penghu University of Science and Technology, 300 Liu-Ho Rd., Makung, Penghu 880, Taiwan, ROC

a r t i c l e

i n f o

Article history: Received 27 April 2011 Accepted 2 March 2012 Available online 13 March 2012 Keywords:

Traffic surveillance system Motion detection Motion estimation Motion compensation Background subtraction Swinging trees filtering Raindrops filtering Shadow elimination

a b s t r a c t

An efficient method for detecting moving vehicles based on the filtering of swinging trees and raindrops is proposed. To extract moving objects from the background, an adaptive background subtraction scheme with a shadow elimination model is used. Swinging trees are removed from foreground objects to reduce the computational complexity of subsequent tracking. Raindrops are removed from foreground objects when necessary. Performance evaluations are carried out using seven real-world traffic image sequences. Experimental results show average recognition rates of 96.83% and 97.20% for swinging trees and rain-drops, respectively, indicating the feasibility of the proposed method.

1. Introduction

The detection of moving vehicles in a video sequence is an important research field due to its applicability to automated visual surveillance systems, traffic monitoring systems, and crime prevention systems. However, existing approaches are affected by luminance variation, weather change, camera jitter, and image noise. For example, some envi-ronmental factors, such as swinging trees and raindrops, in traffic mon-itoring systems greatly impact the tracking performance of vehicles. A robust tracking system is thus desirable.

For a trafﬁc monitoring system, the removal of swinging trees and raindrops is of great importance because they are often erroneously detected as moving vehicles or pedestrians based on background subtraction schemes. Detected objects that are the parts of back-grounds increase the computational burden of the subsequent track-ing of real movtrack-ing objects, such as vehicles and pedestrians, if they are not removed. For most trafﬁc surveillance systems[1,2], the ef-fects of such environmental factors on tracking performance are rarely considered, which limits the systems in some practical appli-cations. The present study thus considers the effects of swinging trees and raindrops on vehicle detection in the proposed system.

For most computer vision-based applications, motion detection that aims to segment regions corresponding to moving objects is of

great importance for the subsequent processes of analysis, recogni-tion, and tracking of those objects. Motion detection can be achieved using techniques which can be roughly classiﬁed as back-ground subtraction[3,4], temporal differencing[5,6], optical ﬂow

[7,8] methods, and block motion estimation [9,10]. Background subtraction can extract the pixels in image sequences with the most discriminative power but it is extremely sensitive to illumination variation. Temporal differencing adapts to dynamic environments but its performance in extracting relevant features is poor. Optical ﬂow can detect moving objects in the presence of camera motion but it is computationally expensive. Block motion estimation can reduce the intensively computational complexity required by the optical-ﬂow method but the accuracy may be also reduced.

Background modeling is of great importance for the detection of moving objects in a video sequence. During the last decade, many different methods for detecting moving objects have been pro-posed[11,12]and several features are used for modeling the back-ground[13]. Stauffer and Grimson[11]modeled each pixel with a mixture of adaptive Gaussians for background estimation to deal with variations in lighting, moving scene clutter, and repetitive motion. In their methods, k-means clustering was used to initially ﬁnd the center of a mixture of Gaussian distributions through an iterative reﬁnement approach. However, k-means may be very slow to converge because it highly depends on the guess of initial center of each cluster. In contrast to a Gaussian mixture model, Elgammal et al. [12] proposed a non-parametric kernel density

http://dx.doi.org/10.1016/j.jvcir.2012.03.002

⇑Corresponding author. Fax: +886 48511245.

E-mail addresses:[email protected](D.-Y. Huang),[email protected]

(C.-H. Chen),[email protected](S.-S. Su),[email protected](W.-C. Hu).

Contents lists available atSciVerse ScienceDirect

J. Vis. Commun. Image R.

(2)

method for background modeling to alleviate the limitations of parametric methods because the assumption of Gaussian distribu-tion for pixel intensities is not always true. In their method, the probability of pixel intensity distribution is calculated directly from the observations without any assumption of the underlying distributions.

Recently, Heikkilä and Pietikäinen[13]have modeled the back-ground using texture features. The features of local binary pattern (LBP) were used in their method due to its tolerance against light-ing changes and its computational efﬁciency. However, the LBP feature cannot work well if the gray scales of the neighboring pix-els are close to the value of the center point. Additionally, many parameters needed to be tuned make this method difﬁcult to be applied to a wide variety of scenarios.

Shadows that can distort the shape of moving objects and thus affect the subsequent task of moving object tracking are a common problem when applying the scheme of background subtraction. To deal with these difﬁculties, Cucchiara et al.[14]proposed a method that uses the features of statistics, adaptivity, and selectivity to de-tect moving objects, ghosts, and shadows. In their method, they adopted color information for both background subtraction and shadow detection to improve the performance of object segmenta-tion and background update. In addisegmenta-tion, weather effect of raining is also an important factor on the performance of moving object detection. Garg and Nayar[15]developed a correlation model to analyze the visual effects of rains on imaging system. Based on the model, they presented an efﬁcient algorithm for detecting and removing rain from video sequences. Moreover, in their exper-iments higher brightness of the raindrops than the corresponding background intensities was also observed.

To consider the color information in background modeling, McKenna et al.[16]used an adaptive background subtraction tech-nique for detecting groups of people by estimating three variance parameters of the R, G, and B channels for each pixel in image se-quences. In their background model, recursive updates are used to adaptively cope with changes in illumination. Background model-ing has also been employed to track movmodel-ing objects. For example, multivariate Gaussian mixtures have been utilized to model pixel color variations for people tracking[17]. Koller et al.[18]used an adaptive background model based on monochromatic images ﬁltered with Gaussian and Gaussian derivative kernels for car tracking.

Liption et al.[5]segmented moving targets from a real-time vi-deo stream using the pixel-wise difference between consecutive frames. Their method can classify humans, vehicles, and back-ground clusters. After classiﬁcation, targets can be tracked by a combination of temporal differencing and template matching. Zhang and Siyal[6]proposed an improved scheme for segmenting moving objects by using two difference images obtained from three consecutive frames in an image sequence, where the two dif-ference images are pre-processed using a logical AND operation.

Garlic and Loncaric[7]adopted optical flow to extract the fea-ture vectors in a video sequence. The extracted feafea-ture vectors are further clustered using a k-means clustering algorithm to deter-mine the characteristic image regions, which allows the detection of moving targets. Gutchess et al.[8]incorporated the concept of optical flow into a background initialization problem, where multi-ple hypotheses of the background value for each pixel are generated by locating periods of stable intensity in image sequences. The like-lihood of each hypothesis is then evaluated using optical flow infor-mation from the neighborhood around a pixel, and the most likely hypothesis is chosen to represent the background.

Background initialization, foreground detection, and back-ground updating are the three important steps for tracking moving objects in an outdoor space. Object tracking methods[19,20]can be classiﬁed as model-based tracking, region-based tracking, active

contour-based tracking, and feature-based tracking. In model-based tracking[21], object tracking is carried out by matching a projected object model to image data, where the object model is produced with prior knowledge. In region-based tracking [16], the variations of image blobs corresponding to moving objects in a video frame are detected to achieve object tracking. Active con-tour-based tracking[22]uses object silhouettes as a bounding con-tour and updates the concon-tour dynamically in consecutive frames. Feature-based tracking [23–25]matches the features of objects, such as color, area, texture, and shape, between successive frames to achieve object tracking.

Schiele[23]used the color of objects with k-means clustering for the extraction and detection of cars as well as people in surveil-lance scenarios using wearable cameras. Their scheme uses neither a priori model of the objects nor a stationary camera; therefore, the results are quite promising. To deal with vehicle tracking in heavy trafﬁc, which may cause parts of vehicles to be occluded, Huang

[26] used features instead of whole contours to track multiple vehicles on freeways. In their scheme, the corner features of a vehi-cle were ﬁrst detected and then tracked by a Kalman ﬁlter. Further-more, to provide an accurate estimate of the vehicle position in each lane, the detection of lane markers is also performed.

In general, most traffic surveillance systems can work well with good weather conditions like sunny day, or with stationary back-grounds that have no repetitive motions such as swinging trees and fluttering flags. It turns out that most existing algorithms fail to detect moving vehicles in scenes containing swinging trees and/or raindrops since these algorithms often erroneously detect swinging trees and/or raindrops as parts of moving objects. The main contribution of this paper is to integrate the modules of filter-ing swfilter-ingfilter-ing trees and raindrops into the background modelfilter-ing to further improve the performance of vehicle detection. The major difference of the proposed method with the existing algorithms

[15,27–29] for removing swinging trees and/or raindrops is the computational efﬁciency of the proposed method because our sys-tem is operated in a real-time outdoor environment.

The rest of this paper is organized as follows. The proposed method of detecting moving vehicles based on the ﬁltering of swinging trees and raindrops is introduced in Section 2. Experi-mental results are provided to demonstrate the performance of the proposed algorithm in Section3. Finally, concluding remarks are given in Section4.

2. Proposed method

Fig. 1shows a flowchart of the proposed method based on the filtering of swinging trees and raindrops for moving object track-ing. Moving objects are first extracted from image sequences using the schemes of motion detection and background updating. Motion detection is used to analyze the temporal correlation of moving ob-jects in successive frames. A frame difference mask and a back-ground subtraction mask are used to acquire the initial object mask using which the problem of stationary objects in back-grounds can be solved. Moreover, boundary refinement is intro-duced to reduce the influence of shadows and the problem of residual backgrounds. The segmented objects are further processed by the proposed modules that remove swinging trees and rain-drops. By removing these undesired moving objects that belong to the background, the performance of moving vehicle detection is significantly improved.

In Section2.1, the method of motion detection for moving ob-jects is described. Background updating is introduced in Section

2.2. Then, the method of shadow removal is explained in Section

2.3. Finally, the methods of removing swinging trees and raindrops are given in Sections2.4 and 2.5, respectively.

(3)

2.1. Moving object segmentation

Trafﬁc monitoring cameras are often installed at ﬁxed sites. The background is thus stationary, making background subtraction

[3,4]suitable for detecting moving objects. Since the performance of segmentation of moving objects highly depends on the extrac-tion of a reliable background, the reference background (or abso-lute background) is set up using a combination of the techniques proposed in[16]to prevent the background information from con-taining moving targets.

To facilitate the process of background extraction, a histogram of gray levels is built for each pixel p(x, y) in a region of the image sequence, as shown inFig. 2(a) and (b). Then, the probability of occurrence, also called appearance probability (AP), for each gray level for a speciﬁc pixel is statistically computed for the given im-age sequence, as shown inFig. 2(c). The different gray levels for a given pixel can be assigned to different classes. For example, 8(=124 117 + 1) classes correspond to gray levels of 117 to 124, respectively, inFig. 2.

Background extraction is then carried out using the AP of each class for a given pixel. For a sequence of video images, the class (or gray level) of a given pixel with the maximum AP value is signed to the background. For instance, the gray level of 122 is as-signed to the background shown inFig. 2(c). Furthermore, each class of a given pixel p(x, y) is assigned a unique number c to iden-tify it. The total number of classes for a given pixel is denoted as nc(x, y). To compute the AP values and classify each gray level, the sum, rc(x, y, c), the class mean,

l

(x, y, c), and the class variance,

r

2_{(x, y, c) of the cth class in a given pixel p(x, y) are used, where the}

sum, rc(x, y, c), records the number of pixels with a given gray level for the cth class, and 0 6 c 6 nc(x, y) 1.

Initialization is carried out for each pixel in the ﬁrst image of a video sequence; at that time, only one class needs to be considered. The frame sampled at time instant t is denoted as f(x, y, t). There-fore, the 0th class is initialized as

l

(x, y, 0) = f(x, y, 0),

r

2

(x, y, 0) = 0, rc(x, y, 0) = 1, and nc(x, y) = 1. Then, a decision function that uses the absolute difference (AD) between the current and mean frames, as shown in Eq.(1), is used to determine whether the gray level of the pixel should be classiﬁed into an existing class or into a new class.

ADðx; y; cÞ ¼ jf ðx; y; tÞ

l

ðx; y; cÞj ð1Þ

The decision function is shown in Eq.(2). The pixel is classiﬁed into class k if the minimum AD(x, y, k) is less than the threshold, Thd. The

parameters of rc(x, y, k),

r

2_{(x, y, k), and}

_l

_{(x, y, k) are then updated}

according to Eq.(3)when AD(x, y, k) is less than Thd; otherwise, a

new class is created according to Eq.(4). In our work, the threshold Thdis set to 2 pixels, indicating that for a given gray level g, the gray

levels of g 1, g, and g + 1 are assigned to the same class.

k ¼ arg min

06c6ncðx;yÞ1ADðx; y; cÞ

pixel class ¼ k if ADðx; y; kÞ < Thd

ncðx; yÞ otherwise 8 > > < > > : ð2Þ

l

ðx; y; kÞ ¼rcðx;y;kÞlðx;y;kÞþf ðx;y;tÞ rcðx;y;kÞþ1

r

2_{ðx; y; kÞ ¼} 1 rcðx;y;kÞþ1 ½rcðx; y; kÞ

r

2ðx; y; kÞ þ j

l

ðx; y; kÞ f ðx; y; tÞj 2 n o rcðx; y; kÞ ¼ rcðx; y; kÞ þ 1 8 > > < > > : ð3Þ

l

ðx; y; ncðx; yÞÞ ¼ f ðx; y; tÞ

r

2_{ðx; y; ncðx; yÞÞ ¼ 0} rcðx; y; ncðx; yÞÞ ¼ 1 ncðx; yÞ ¼ ncðx; yÞ þ 1 8 > > > < > > > : ð4Þ

As image frames are collected over time, the class (or gray level) of each pixel assigned to the background is established. The AP(x,y,c) of the cth class of a pixel point is calculated using Eq. (5). The mth class that has the maximum AP value is then assigned to the background, determined using Eq.(6).

Fig. 1. Flowchart of the proposed method for moving object tracking.

Fig. 2. Illustration of the appearance probability of each class for a given pixel. (a) Pixel p(x, y) at a location in a sequence of video frames, (b) histogram of gray levels for a ﬁxed point in consecutive frames, and (c) appearance probability for each gray level in (b).

(4)

APðx; y; cÞ ¼_P_ncðx;yÞ1rcðx; y; cÞ

c¼0 rcðx; y; cÞ

ð5Þ

m ¼ arg max

06c6ncðx;yÞ1APðx; y; cÞ ð6Þ

The background model is then set up as:

Bðx; yÞ ¼

l

ðx; y; mÞ

r

2_{ðx; yÞ ¼}

_r

2_{ðx; y; mÞ}

ð7Þ

where B(x, y) is established from the class m that has the maximum AP value of a given pixel p(x, y), and

r

2_{(x, y) is the variance of the}

background. The detection function of background subtraction is:

ifjf ðx; yÞ Bðx; yÞj >

x

r

ðx; yÞ f ðx; yÞ 2 foreground pixel else

f ðx; yÞ 2 background pixel

ð8Þ

where

x

is the weight assigned to the frame difference threshold with a range of 1 6

x

6_5.

Because pixels with similar gray levels can exist in both the foreground and background, several holes might be generated in the binary mask of moving objects. To solve this problem, erosion and dilation of morphological operations and the fast 8-connected component labeling method[30]are used. These methods also re-move the noise in the binary image. A complete mask of the mov-ing object is thus obtained for better extraction. The captured image sequence may contain multiple moving objects, so a simple multi-object segmentation algorithm[31] is employed to extract them.Fig. 3shows the results of the proposed method for the seg-mentation of a moving object.

2.2. Background updating

The basic idea of background updating is based on the fact that the background changes with time. If the background is not up-dated, the performance of motion detection will decrease. There-fore, an iterative rule is proposed here to carry out the background updating. The background is updated by taking a weighted average of the current background and the current frame for a sequence of video images. To extract background pixels for updating the background of the current frame, the updating rule is:

Bðx; y; t þ 1Þ ¼ Bðx; y; tÞ; if f ðx; y; tÞ 2 foreground ð1

a

ÞBðx; y; tÞ þ

a

f ðx; y; tÞ; if f ðx; y; tÞ 2 background 8 > < > : ð9Þ

r

2 ðx; y; t þ 1Þ ¼

r

2_{ðx; y; tÞ;} _{if f ðx; y; tÞ 2 foreground} ð1

a

Þ

r

2_{ðx; y; tÞ þ}

_a

_{ðf ðx; y; tÞ Bðx; y; tÞÞ}2 ; if f ðx; y; tÞ 2 background 8 > < > : ð10Þ

where

a

2 [0, 1] denotes the weight (or updating rate) assigned to the current frame and background. In this work, the optimal value of

a

is 0.05, as obtained from experiments.

2.3. Shadow removal

Shadow removal is the important part of motion detection. Seg-mented moving objects containing shadows are commonly ob-tained when using a method of motion detection based on the background subtraction technique. Shadow pixels are primarily caused by light ﬂuctuation, which may lead to the deformation of segmented objects, or worse, falsely connected objects. The sha-dow elimination model described below is thus proposed.

Let YFn(x, y) and YBn(x, y) be the light intensities of the current

frame and its background, respectively, at a pixel p(x, y) for the nth frame in an image sequence. Hence, the ratio of light intensity, BRn(x, y), can be represented as:

BRnðx; yÞ ¼ YFnðx; yÞ=YBnðx; yÞ;

Illuminated pixels; if BRðx; yÞ > 1 Shadowed pixels; if BRðx; yÞ < 1

ð11Þ

where BRn(x, y) > 1 represents illuminated pixels detected in

seg-mented objects for the current frame and shadow pixels for BRn(x, y) < 1 otherwise. Using the ratio of the light intensity of the

current frame to that of the background, the proposed shadow elim-ination model is given as:

SPnðx; yÞ ¼

1; if ð

a

6_BR_n_{ðx; yÞ 6 b and}

jSFnðx; yÞ SBnðx; yÞj 6

s

sand

jHFnðx; yÞ HBnðx; yÞj 6 ð

s

HÞ 0; otherwise 8 > > > < > > > : ð12Þ

Shadow pixels are detected when SPn(x, y) = 1; otherwise, they are

pixels belonging to moving objects. To allow some noise in the background image, the value of b in our work is taken as 0.95. We also observed that when the value of BRn(x, y) is too low, those

pix-els may belong to moving objects rather than the parts of shadow; therefore, we set the value of

a

to be 0.65 by our experiments. In Eq.

(12), SFn(x, y) and SBn(x, y) represent the saturation values in the HSV

color space for the nth frame and its background, respectively, and HFn(x, y) and HBn(x, y) denote their corresponding hue values. As

ob-served from experiments, the differences in values of saturation and hue between the current frame and its background are commonly not large; therefore, the values of

s

sand

s

Hin our work are taken

as 0.6 and 70, respectively. The resulting shadow pixels (in red) de-tected by the proposed algorithm are shown inFig. 4.

2.4. Removal of swinging tree leaves

Most existing algorithms used in trafﬁc monitoring systems of-ten erroneously detect swinging trees as moving vehicles or pedes-trians, particularly when surveillance cameras are installed in an outdoor space. This leads to a heavy computational burden for sub-sequent tracking procedures. An effective algorithm for removing swinging trees is thus proposed here. The algorithm ﬂowchart is shown inFig. 5and the results are shown inFig. 6.

2.4.1. Motion estimation and motion compensation

The algorithm of ﬁltering swinging trees is as follows. First, mo-tion estimamo-tion is carried out using the difference frame (DF) and the corresponding background (BF) based on the fact that the moving objects due to swinging trees in the current frame can be estimated from its background. In this paper, the difference frame and its background frame are divided into blocks with W(width) H (height) pixels, where W H is the size of the tem-plate block in the difference frame shown inFig. 7(a), and W and H are both set to 2 pixels. To achieve the optimal matching of the template block in the background frame, the cost function of the mean absolute error (MAE) given in Eq.(13)is evaluated, where vector (x, y) is the candidate motion vector, and DF(, ) and BF(, ) indicate the template block in the difference frame and the match-ing block in the background frame, respectively.

MAEðx; yÞ ¼ 1 W H P W1 i¼0 P H1 j¼0 jBFðx þ i; y þ jÞ DFði; jÞj ð13Þ

Motion compensation is then achieved using the full search algo-rithm within the speciﬁed search region given in Eq.(14), where the search parameters

a

and b are both set to 3. The best motion

(5)

vector (xbest, ybest) can be obtained using Eq.(15)by scanning all the

candidate blocks within the search region in the background frame shown inFig. 7(b).

RS¼

a

W bH ð14Þ

ðxbest;ybestÞ ¼ arg min ðx;yÞ2RS

MAEðx; yÞ ð15Þ

The block with the best motion vector in the background frame is then used to replace the template block in the difference frame.

The resulting motion compensation of the difference frame (see

Fig. 6(c)) is shown inFig. 6(d).

2.4.2. Filtering of swinging tree leaves

The ﬁltering of swinging trees is carried out using the informa-tion of both the moinforma-tion compensainforma-tion frame (MCF) and the differ-ence frame (DF). To avoid the block effects of both MCF and DF, they are preprocessed by a median ﬁlter. Even though the shape

Fig. 3. Results of the proposed method for the segmentation of a moving object. (a) Input image, (b) background image, (c) binary difference image between (a) and (b), and (d) results of performing both the morphological and the fast 8-connected component labeling operations.

Fig. 4. Results of shadow pixel detection using the proposed shadow elimination model. (a and c) Original images and (b and d) pixels of shadows (red) and moving scooter (blue). (For interpretation of the references to colour in this ﬁgure legend, the reader is referred to the web version of this article.)

(6)

of swinging trees is irregular, they can be classiﬁed into two cate-gories for the three test image sequences used here in terms of the compactness (i.e., Rc = perimeter2/area) of the segmented objects as type-I (seeFig. 8(d)) if Rc < Tc, and type-II (seeFig. 8(e) and (f)) if Rc P Tc, where Tc 2 [0, 1] is a threshold (set to 0.3, as obtained from experiments). The ﬁltering model of swinging trees for type-I is proposed as:

TObjkðIÞ¼

1; if ðYmObjk 6Yt1Þ and ðHmObjk6Ht1Þ

0; Otherwise ð16Þ where YmObjk ¼ 1 NObjk P N_Objk1 i¼0

jYObjkðDFÞðiÞ YObjkðMCFÞðiÞj ð17Þ

and HmObjk ¼ 1 NObjk P N_Objk1 i¼0

jHObjkðDFÞðiÞ HObjkðMCFÞðiÞj ð18Þ

where Y and H denote the gray intensity and hue value in the HSV color space, respectively. YmObjk and HmObjk represent the mean

difference value of gray intensity and hue value of the kth object estimated from both the difference frame and the motion compen-sation frame, respectively. YObjkðDFÞðiÞ and YObjkðMCFÞðiÞðHObjkðDFÞðiÞ and

HObjkðMCFÞðiÞÞ are the gray intensities (hue values) at pixel i of the

Fig. 5. Proposed algorithm for removing swinging trees based on motion estima-tion and moestima-tion compensaestima-tion.

Fig. 6. Results of ﬁltering swinging trees using the proposed algorithm. (a) Original image; (b) absolute background of (a); (c) difference frame between (a) and (b); (d) results of motion compensation of (c); (e) results of ﬁltering swinging trees.

(7)

kth object in the difference frame (DF) and the motion compensa-tion frame (MCF), respectively. NObjk is the total number of pixels

in the kth object. Yt1and Ht1are the thresholds (set to 30 and 90,

as obtained from experiments), respectively. TObjkðIÞ¼ 1 denotes that

the kth object is type-I swinging trees that should be ﬁltered out from the foreground objects, as shown inFig. 6(e).

Swinging trees are categorized as type-II when the criterion of Rc P Tc is met. For this case, the ﬁltering model is proposed in Eq. (19), where YmObjk and HmObjk were previously deﬁned for

Eqs. (17) and (18), respectively. The thresholds of Yt2 and Ht2 in

Eq.(19)are set to 40 and 120, respectively. TObjkðIIÞ¼ 1 represents

the kth object detected as type-II swinging trees. The results of removing swinging trees of type-I and type-II are shown in

Fig. 8(g), (h), and (i).

TObjkðIIÞ¼

1; if ðYmObjk6Yt2Þ and ðHmObjk6Ht2Þ

0; Otherwise

ð19Þ

2.5. Removal of raindrops

In an outdoor space, raindrops are often falsely detected as moving objects. Hence, raindrops should be removed. Raindrop fil-tering reduces the computational complexity of subsequent object tracking. An effective and efficient algorithm, shown inFig. 9, is thus proposed to filter out raindrops. In the proposed method, the matching of the gray intensity feature, appearance feature, and temporal feature is first carried out to identify raindrops.

Raindrops have three salient features: a gray intensity feature, an appearance feature, and a temporal feature. Thus, moving objects can be identiﬁed as raindrops according to the three features. (1) Gray intensity: raindrops have a higher gray intensity than those of the corresponding pixels in the background. (2) Appearance:

rain-drops appear as a long strip. (3) Temporal: rainrain-drops in consecutive frames do not overlap when the frame rate is 30 frames per second (FPS) (FPS for the video cameras used here).

2.5.1. Gray intensity feature of raindrops

To illustrate the gray intensity feature of raindrops, two image frames were presented. One is the difference frame (DF) shown inFig. 10(c), which was obtained by subtracting the background from the original image. The other is the difference positive frame (DPF) shown inFig. 10(d), which was obtained from the DF frame but only extracts the pixels that have higher gray intensity in the

Fig. 8. Results of removing of swinging trees for three test image sequences. (a, b, and c) Original images for the three test image sequences; (d, e, and f) segmented moving objects in (a), (b), and (c), respectively; (g, h, and i) results of removing of swinging trees.

(8)

[30] D.Y. Huang, C.J. Lin, W.C. Hu, Learning-based face detection by adaptive switching of skin color models and AdaBoost under varying illumination, Journal of Information Hiding and Multimedia Signal Processing (JIHMSP) 2 (3) (2011) 204–216.

[31] T.Y. Chen, C.H. Chen, D.J. Wang, A cost-effective people-counter for passing through a gate based on image processing, International Journal of Innovative Computing, Information and Control (IJICIC) 5 (3) (2009) 785–800.

[32] Database for background modeling. Available from: <http://perception.i2r.a-star.edu.sg/>.

[33] T.Y. Chen, C.H. Chen, D.J. Wang, Y.C. Chiou, Real-time video object segmentation algorithm based on change detection and background updating, International Journal of Innovative Computing, Information and Control (IJICIC) 5 (7) (2009) 1797–1810.