A Moving Objects Detection Algorithm using Iterative ... - IEEE Xplore

0 downloads 0 Views 317KB Size Report
Iterative Division and Gaussian Mixture Model. Fei He, Jin Wang, Xuejie Zhang, Yun Gao. School of Information Science and Technology. Yunnan University.
A Moving Objects Detection Algorithm using Iterative Division and Gaussian Mixture Model Fei He, Jin Wang, Xuejie Zhang, Yun Gao School of Information Science and Technology Yunnan University Kunming, China [email protected], [email protected], [email protected], [email protected]

proposed many improvement methods for GMM, but result is still non-ideal [5, 6, 7]. Bourezak [8] presents an algorithm for the detection and tracking of moving objects based on color and texture analysis for real time processing. This moving objects detection algorithm is based on color histograms and iteratively divided interest regions for motion detection. The main idea is to split empirically the image into squared regions of the same size W*H. At each step, color histograms of both reference frame and the current frame are calculated for each region. Then, the distance metric is used to measure the distance between both histograms. If the distance is larger than a specified threshold, then the regions are different, and motion is detected in that region. For each one motion detected region is split in four smaller squared region, that is W’=W/2, H’=H/2. Then splitting and histogram distance calculating will be repeated until the region size equals a specified threshold. Iterative division has definite effects anti-illumination variation, noise and changes in the scene such as tree leaves motion et al. However, it still has some shortcomings: lack of background update method; iterative more times will lower algorithm performance; cannot get the clearly contour of moving objects. In case of focusing on static camera, this paper presents an algorithm for moving objects detection using iterative division and Gaussian mixture model. First, we choice the first frame as background frame and build GMM on each pixel. Second, we use iterative division algorithm for comparing current frame with background frame to obtain motion region. In this step, iterate only 1 or 2 times. In order to obtain moving objects region as much as possible, in each iterative, region’s threshold will be automatic selected. Third, we compare each pixel in motion region with same position pixel in background frame using GMM. In this step, we use GMM to obtain the pixels of moving objects and update background frame. Pure GMM is sensitive to local variations, noise and illumination variation. Pure iterative division is low performancelack of background update method and bad contour of moving objects. To tackle this issue, we propose an algorithm that combines GMM and iterative division. The experimental results indicate that our

Abstract—In the fields of intelligent video surveillance, video analysis and computer vision, moving objects detection is one of the fundamental tasks. In this paper, we present an algorithm for moving objects detection based on Gaussian mixture model and iterative division, which can be used for real time processing. We use iterative division algorithm to filter illumination variation and local variations. Then, GMM is used to update background and get pixel level precision. Experimental results show that our method, by the combination of iterative division and GMM, is able to overcome deficiency of tradition method and gain good detection results. Keywords: motion detection; background subtracktion; iterative division; Gaussian mixture model

I. INTRODUCTION In the fields of intelligent video surveillance, video analysis, computer vision, moving objects detection is one of the fundamental tasks. The higher levels such as moving object classification, tracking and behavior understanding heavily depend on the results of moving objects detection. Nowadays, focusing on a video sequences from static camera, there are a lot of methods for moving objects detection. The most important methods list in [1, 2]. We note that most existing approaches are based on pixel comparison, and that all the pixels are independent. Background modeling based on Gaussian mixture model (GMM) [3, 4] is one of the most successful moving objects detection algorithms. GMM allows moving objects in background; in moving objects detection process, GMM could dynamic update background; in case of no illumination variation and no local variations, GMM could obtain very good foreground. However, the deficiency of GMM is very clear. GMM requires a background learning process; in a busy scene, when moving object is big and slowly, the result of detection is non-ideal; GMM is sensitivity to illumination variation; if there is perturbation in background (e.g. shaking trees or swaying curtain), result of detection is non-ideal. Recently, many researchers have

978-1-4244-5848-6/10/$26.00 ©2010 IEEE

229

algorithm is better than pure GMM and pure iterative division. II.

APPROACH

A.

Preprocessing Video capture is done in the RGB color space; however this color space is not suited for our application because small changes in the light intensity change significantly an object description. We prefer a color space less sensitive to light intensity. Thus, we choice the HSV color space for our algorithm because it conforms to color cluster feature of human's vision [9]. The HSV color space provides a direct handle to control chromaticity present in Hue and Saturation. We should give H and S handles more importance in the quantization phase. Nonetheless, Hue is considered to be more reliable for color segmentation. So, the H, S, V handles are quantized in 162 bins (18X3X3). Before converting from RGB to HSV space, a 3X3 median filter is applied to clean the image from acquisition noise. B. Histogram matching Intuitively, image is formed by many color lumps. We can split empirically the image into squared regions. When a color lump is moving in a region, histogram of the region will not be changed. In case of illumination variation, histogram will be shifted, but shape of the histogram and distribution of the histogram changes indistinctively. So using histogram as a feature is feasible. In this paper, the L1 distance metric and the L2 distance metric are used to measure the distance between both histograms. The L1 distance metric is defined as

naturally for partial matching. It is more robust than histogram matching techniques, in that it can operate on variable-length representations of the distributions which avoid quantization and other binning problems typical of histograms. When used to compare distributions with the same overall mass, the EMD is a true metric. EMD can be formalized as the following linear programming problem: Let P={(p1,wp1),…,(pm,wpm)} be the first signature with m clusters, where pi is the cluster representative and wpi is the weight of the cluster; let Q={(q1,wq1),…,(qn,wqn)} be the second signature with n clusters; and D=[dij] the ground distance matrix where dij is the ground distance between clusters pi and qj; F=[fij] is a flow, with fij the flow between pi and qj, that minimizes the overall cost m

n

WORK ( P, Q ) = ∑ ∑ d ij f ij i =1 j =1

(2)

subject to the following constraints fij ≥ 0, i = 1...N , j = 1...N m

∑f

ij

≤ wpi , i = 1...m

i =1

(4)

n

∑f

ij

≤ w pj , j = 1...n

j =1

m

n

∑∑ f i =1 j =1

ij

(3)

(5) m

n

i =1

j =1

= min( ∑ w pi , ∑ wqj )

(6) Then the distance between signature P and signature Q can be defined as WORK ( P , Q) EMD( P, Q) = m n ∑ i =1 ∑ j =1 fij (7) EMD essentially measures how much work it would take to “shovel” one histogram shape into another, including moving part (or all) of the histogram to a new location. Let L2=EMD(hist1, hist2), L2 ∈ [0,1] . If L2 is less than a specified threshold t2, then two histograms have a similar shape. If L2 is larger than a specified threshold t2, motion is detected in that region.

162

L1 = ∑ | H ref (i ) − H cur (i ) | i =1

(1) In (1), Href is current region’s histogram in background frame, Hcur is the same region’s histogram in current frame, H(i) is the ith color’s value in histogram. Href and Hcur will be normalized. L1 returns the sum of the levels of two histograms, L1 ∈ [0, 2] . If L1 is larger than a specified threshold t1, then the regions are different, and motion is detected in that region. The threshold is fixed as a percentage of the square size, according to the level of change we want to detect. That is, the smaller is the square size, the larger will be the value of the threshold. Illumination variation can cause shifts in color values, although such shifts tend not to change the shape of the histogram of color values, but shift the color value locations and thus cause the L1 histogram-matching fail. However, illumination variation causes region’s histogram shifting. L1 cannot reflect this change. We use earth mover’s distance algorithm to modify the result [10, 11]. Earth mover’s distance (EMD) is based on a solution to the transportation problem from linear optimization, for which efficient algorithms are available, and also allows

C. Iterative Division For more specifically, we define current frame is F and background frame is B, their width and height are W and H. We split F and W into N regions of same size w*h, N = W/w*H/h. Iterative division motion detection algorithm steps are Step1. F and B are first split into N regions of size w*h, the value of w and the value depend on the size of the objects. The larger the object is relatively to the frame, the larger are w and h. Regions are marked F(i) or B(i), i ∈ N . Step2. For each region, we calculate L1 and L2 between each pair of histograms. If L1