APP下载

Background modeling based on improved parameter estimation for moving target detection

2018-07-24YiqiangZHUXilongDING

机床与液压 2018年12期

Yi-qiang ZHU, Xi-long DING

(1School of Electric Engineering, Wuhu Institute of Technology, Wuhu 241006, China) (2Weifang University of Science & Technology, Sino-India Computer SoftWare Institute, Shouguang 262700, China)

Abstract: In order to improve the recognition rate of target detection, a background modeling algorithm based on improved Student-t distribution parameter estimation is proposed in this paper, which is suitable for all kinds of motion video target detection such as sports video application field. Firstly, the background modeling method based on the traditional finite model is analyzed. Then, the Student-t distribution is used to construct the background model parameter estimation method. The method is used to complete the parameter estimation by improving the maximum expectation algorithm, and the parameter space is segmented. The simulation results show that the segmentation effect of the proposed algorithm is better as compared with the Gaussian background modeling, K-means and fuzzy C-means clustering background modeling, and the motion target detection accuracy is higher and the training speed is faster. The results verify that the validity and advancement of the proposed algorithm.

Key words: Moving object detection, Student-t, Background modeling, Background segmentation

In real life, with the continuous development of computer vision technology, various new applications appear constantly, such as face recognition, sports human target detection and tracking, driving auxiliary, robot control and video surveillance [1-3]. The main source of information for people is images, where video monitoring is necessary for many important activities, such as enhancing public safety or protecting important infrastructure. The falling cost of high-quality cameras has made the video monitoring system more popular and more affordable. This growth creates a huge demand for real-time video processing, especially real-time detection and tracking [4].

Moving target detection and tracking technology has always been an important research area of people’s concern [5]. At present, there are three kinds of main motion target detection methods [6]: ① optical flow method; ② inter-frame difference method;③ background subtraction division. The accuracy of optical flow method is higher, but the operation complexity is larger, and the sensitivity of light is higher, so the scope of application becomes narrower. The denoising ability of the frame difference method is better, but the degree of dependence on video sequence is higher, and the detection accuracy will be reduced when the movement speed gets changed. Due to the good applicability, the background subtraction method has been widely concerned by the majority of researchers and has become the mainstream of moving target detection technology. The background subtracting method can be used to break up the movement objects well and satisfy the demand of most video monitoring scenarios. The accuracy of background subtracting method will not be influenced by the motion speed, and its detection performance is directly determined by the performance of the background modeling. In real application scenarios, the effects of light change, background interference, shadow interference and camera shake will have a bad effect on the detection of moving targets.

The background clustering based on probabilistic statistics is the most popular method at present, which could be mainly divided into two kinds: parameters and non-parameters. The parametric literature [7] and literature [8] have proposed an improved hybrid Gaussian model parameter estimation method based on probability statistics and applied to different image processing scenarios. The non-parametric method based on k mean is proposed in literature [9]. Although non-parametric methods are better in their applicability, they can only be used for a shorter time series of video, therefore, most studies are more likely to use parameterized methods.

A Gaussian mixture model based on spatial correlation is proposed in reference [10]. By optimizing the parameters with the gradient descent method, the complexity of the algorithm could be reduced, and the influence of noise on image segmentation will be effectively removed. However, in practical application, the probability distribution of pixels in video image sequence is not consistent with the Gaussian distribution due to the variation of illumination and camera jitter. Student-t distribution has better robustness in the large noise test data set because of its statistical characteristics with spikes and it can improve the detection accuracy of image recognition applications [11].

Therefore, this paper proposed a background modeling algorithm based on the improvement parameters of the Student-t distribution estimation. Firstly, the background modeling method based on finite mixture model is analyzed. Then, the parameter estimation method of background model is constructed by using the Student-t distribution. This method is used to estimate the parameters by improving the expected value, and the parameter space will be segmented. As compared with the Gaussian background modeling, K-means and fuzzy C-means clustering background modeling, the segmentation effect of the proposed algorithm is better, and the motion target detection accuracy is higher and the training speed is faster, all the results verifies the validity and advancement of the proposed algorithm.

1 Background modeling based on the finite mixture model

1.1 Parameter estimation of finite mixture model

Since the single probability distribution only exists in single peak, the probability distribution of most actual application scenarios is composed of multiple peaks. Therefore, for the actual application scenario, the probability distribution of multiple peaks should be used to simulate the simulation effectively. The finite mixed model can simulate the data set better, and obtain the linear superposition of the probability distribution function of each cluster by statistical method. The formula for the finite mixed model could be defined as follows:

(1)

1.2 Gaussian mixture model

For formula (1), ifP(x|θk) is a Gaussian distribution, thenf(x) is the Gaussian mixture model. Letxidenote the i-th pixel of the image,i=1,2,…,N. The probability density function of the current pixel has a linear combination of class K, andΩjis the j-th class,j=1,2,…,K. The probability density function ofxicould be expressed as:

(2)

(3)

Where,μjis the mean of the Gaussian distribution andσjis the covariance of the Gaussian distribution. According to the formula (2), the logarithmic likelihood function could be obtained as follows:

(4)

The posterior probability of the i-th pixel can be expressed as:

(5)

Generally speaking, use the empirical solution to achieve the log likelihood function:

(6)

In order to obtain the largest log likelihood function, the formula (6) must be minimized:

ln(πijpnew(xi|Ωj))

(7)

Where,εdenotes the error function. In order to obtain the largest log likelihood function,εmust be minimized and the optimized parameters are given to equation (5).

2 Proposed Student-t distribution background modeling

2.1 Student-t distribution model

We setX={X1,…,Xi,…,XN} to represent the independent distribution data with dimension d, which is divided into K types. The Student-t distribution can be expressed ast(μk,Ηk,vk). Where,vis the degree of freedom of the Student-t distribution,μis the mean of the Student-t distribution, andΗkis the symmetric and positive definite matrix. The smaller the value ofv, the longer the tail length of the Student-t distribution as shown in Fig.1. When the degree of freedomvis infinite, the variance of Student-t distribution and Gauss distribution are the same.

The Student-t distribution model of each pixel in the image sequence is as follows:

(8)

Where,S(Xi,μk,Hk,vk) denotes the probability density function of component t in a cluster, and it could be calculated as follows:

(9)

Where,Γ(·) denotes a gamma function.

(10)

Fig.1 the influence of different v on Student-t distribution

2.2 Improved model parameter estimation

In order to improve the convergence rate, this paper presents an improved maximum expectation algorithm to complete the model parameter estimation. The method sets thatΘrepresents a set of parameters for a finite mixed model andP(Θ) represents the prior probabilities of the parameters. DivideΘinto two parts {Θ1,Θ2}, where,Θ1={πk,μk,Hk} andΘ2={vk}. Firstly, theμikis updated by usingt+1-thΘ1, and then the degree of freedom could be estimated after repeated m times.

The concrete steps are as follows:

(11)

(12)

(13)

(14)

(15)

(4) Repeat execution (3)~(5) steps, repeat m times.

Φ(0.5vk)+Φ(0.5(vk+p))-

log(0.5(vk+p))=0

(16)

(17)

3 Experiment and comparison

We use the MATLAB programming language to implement the improved Student-t distribution parameter estimation algorithm, and compare it with the Gaussian Mixture Model (GMM) [11] and fuzzy C-means (FCM) clustering background modeling [13]. Experimental environment parameters are as follows: 2.4 GHz dual-core CPU, 4G memory, simulation environment for the Matlab R2014a.

3.1 Target test results

In order to make the subjective evaluation of the proposed algorithm, the experimental test uses the Change Detection 2014 image sequence data set [14], and the sequence image size is 320×240. The target detection results of three different methods are used for video sequence 1, as shown in Fig.2 (c), (d) and (e). The target detection results of the three different methods are used for the video sequence 2, as shown in Fig.3 (c), (d) and (e). It can be seen that the algorithm presented in this paper is superior to other methods in terms of background segmentation performance.

Fig.2 Target test on instance 1

Fig.3 Target test on instance 2

As shown in Fig.2 and Fig.3, it can be seen from the intuitive point that the background segmentation effect of proposed algorithm is the best. As compared with the other two algorithms, it is possible to extract the detected target completely under the motion scene, which is helpful to improve the accuracy of the moving target detection.

3.2 Performance evaluation

In order to quantitatively evaluate the performance of motion target detection, the background detection error (B_error), target detection error (F_error) and overall recognition rate (WRR) are used as objective parameters to evaluate the performance of the algorithm[8-9].

(18)

(19)

(20)

Where,Brepresents the total number of pixels in the background inside the ground solution,Frepresents the total number of pixels in the foreground area in the ground solution,BNindicates that when the background area is erroneously detected as a moving object,FNindicates that when the target area is erroneously detected as Background,BPorFPmeans correct detection. The smaller the value ofB_error andF_error, the better the value ofWRR.

After average the results of video series 1 and video sequence 2, the comparison results of the detection rates of the three different methods are shown in Table 1 and Fig.4. It can be seen from Table 1 and Fig.4 that the proposed algorithm has the lowest background false detection rate and target false detection rate (i.e., 52.21% and 0.98), and has the highest overall recognition rate (i.e., 63.71).

Table 1 Comparison results of detection rate

B_error/%F_error/% WRR/% FCM54.326.4757.61GMM 53.143.2162.05Ours52.210.9863.71

Fig.4 Comparison results of detection rate

The detection results after morphological opening and closing operations are shown in Table 2. The most optimal values of the algorithm are presented, in which the background false detection rate and the false detection rate are 0.07% and 2.39%, respectively, and the overall recognition rate is 94.43%.

Table 2 Detection results after morphological opening and closing operation

B_error/% F_error/% WRR/% FCM0.0916.4881.56GMM 0.089.7287.37Ours0.072.3994.43

3.3 Algorithmic complexity analysis

In order to analyze the efficiency of the algorithm, the complexity of the proposed algorithm is evaluated. The training time is calculated when the three different algorithms are used to construct the background model. In the tests, the training time of each pixel in the image will be calculated. Finally, the training time of all the pixels is simply summed. The comparison results of the training time for each pixel are shown in Table 3.

Table 3 Comparison of algorithm processing speed in seconds

Video sequence 1Video sequence 2FCM0.0180.021GMM 0.0470.051Ours0.0860.091

As can be seen from Table 3, because of the iterative calculation process for the degree of freedom in the proposed algorithm, the steps will get increased to build the model, so the training speed of each pixel have been reduced as compared to the other two methods. The subsequent study will focus on solving the above-mentioned problem.

4 Conclusion

In this paper, we propose a background modeling algorithm based on improved Student-t distribution parameter estimation. Firstly, the background modeling method based on the traditional finite model is analyzed. Then, the Student-t distribution is used to construct the background model parameter estimation method. The method is used to complete the parameter estimation by improving the maximum expectation algorithm, and the parameter space is segmented. The simulation results show that the segmentation effect of the proposed algorithm is better as compared with the Gaussian background modeling,K-means and fuzzyC-means clustering background modeling, so the motion target detection accuracy is higher and the training speed is faster, and the validity and advancement of the proposed algorithm has been confirmed.


登录APP查看全文