CN104599290B - Video sensing node-oriented target detection method - Google Patents

Abstract

The invention discloses a target detection method oriented to video sensing nodes, including compressed sensing steps, background modeling steps, update steps and post-processing steps, setting different sampling values M according to the interest area, and the target block 1.2 in the previous frame Double the area to increase the sampling rate, and reduce the sampling rate in the background area; and, when the background brightness changes small, reduce the number of Gaussian distributions modeled to reduce the learning rate; when the brightness changes greatly, increase the number of Gaussian distributions to reduce the learning rate. Increase the learning rate.

Description

An Object Detection Method Oriented to Video Sensing Nodes

technical field

The invention relates to the technical field of target detection, in particular to a target detection method for video perception nodes.

Background technique

The wireless video sensor network is an "intelligent" autonomous measurement and control wireless network system composed of a large number of video nodes with communication and computing capabilities arranged in a specific way or randomly in the monitoring area. The video sensor nodes have a strong collaborative ability, and complete the global task through local image data collection, processing and data interaction between nodes. Compared with the traditional monitoring mode, the use of wireless video sensor networks to build a distributed intelligent monitoring system has the advantages of unattended, wide coverage, stable performance, high flexibility, and any combination of monitoring scenarios, especially suitable for traffic intersections, airports, etc. Target tracking and event monitoring in critical areas such as subway stations or harsh environments.

In computer vision and wireless video sensor network-related applications, the detection of objects in acquired video images is the first step. The quality of the target detection algorithm affects further visual processing such as follow-up tracking and behavior recognition. Due to the complexity and changeability of the actual scene, the existing target detection algorithms are generally complex, with a large amount of calculation and high memory capacity requirements, and are not suitable for video perception nodes with limited resources.

Therefore, the target detection algorithm for video perception nodes must first consider the problem of algorithm efficiency, and reduce the amount of calculation and storage capacity as much as possible. The compressed sensing theory breaks through the requirement of the number of samples under the traditional Laquist theory. As long as the signal is compressible or sparse, the transformed high-dimensional signal can be sampled through an observation matrix that meets certain conditions to obtain a sampled low-dimensional signal. Then solving an optimization problem can perfectly reconstruct the original signal from a small number of sampled values. Applying the compressive sensing theory to the target detection algorithm based on the background subtraction method can greatly reduce the number of pixels involved in the background modeling while retaining the original image information, thereby improving the efficiency of the algorithm. Therefore, on the basis of studying several commonly used background modeling methods today, a Structured Compressive Sensing Adaptive Gaussian Mixture Model (SCS-AGMM) background modeling algorithm based on structured compressed sensing is proposed to construct the structure The stochastic measurement matrix is used to reduce the amount of data involved in background modeling, and the performance of the algorithm is optimized from many aspects.

The background subtraction method is a relatively mature method in the field of target detection and is widely used. This method subtracts the pixel value of the current frame of the video image from the corresponding position of the background model, and when the absolute value of the difference is greater than a certain threshold, it is determined that the pixel is a target pixel, otherwise it is a background pixel. And through post-image processing, a complete target image is obtained.

For complex and dynamically changing backgrounds, such as fluctuating water surfaces, shaking trees, trembling cameras, etc. in the scene, the probability density distribution map of pixel values often shows a bimodal or multimodal state. This requires the use of a linear combination of multiple Gaussian distributions to accurately model the background, which is called a Gaussian mixture model (GMM). Using GMM to build a background model for each pixel in the image can adapt to the changes in illumination and the interference of moving backgrounds in video images.

In recent years, a large number of improved algorithms based on mixed Gaussian background modeling have emerged. The advantages of these methods are that they have better detection effects and can remove motion interference in complex background situations; the disadvantages are that they have a large amount of calculation and storage, and the running speed is slow, so they are not suitable for Resource-constrained video-aware nodes.

Contents of the invention

Aiming at the above-mentioned deficiencies in the prior art, the present invention provides a target detection method oriented to video perception nodes. The present invention is realized through the following technical solutions:

A target detection method for video perception nodes, comprising steps:

Image reconstruction step: divide the image into blocks according to the size of the collected images, and convert the collected image blocks into N×1 vectors;

Compressive sensing step: construct a structured random measurement matrix to sample and compress the converted vector;

Background modeling step: use the adaptive mixed Gaussian model to perform Gaussian modeling on each measured matrix block, and use the least pixel rule to detect the target block and background block; use the mixed Gaussian model to establish at least one pixel for each pixel of each candidate node. Background model, initialize the background model with the first frame of image data, set a uniform background threshold for each background model, the background model whose pixel weight is greater than the background threshold describes the background distribution, and the pixel weight is less than or equal to The background model of the background threshold describes the foreground distribution. Use the pixels judged as the background in the image to reinitialize the background model whose weight is less than the initialization threshold; Matching detection, determine that the pixel points whose background models do not match the current pixel values are points in the target area, update the distribution parameters for the successfully matched background models, and update the weights for each background model;

Update step: use different strategies to update the target block and background block, and adjust the parameters of the structured random measurement matrix according to the detection results;

Post-processing step: performing post-processing on the detected target image to obtain the final target image;

Among them, different strategies include setting different sampling values M according to the region of interest, increasing the sampling rate in the area of 1.2 times the target block in the previous frame, and reducing the sampling rate in the background area; and, when the background brightness changes little, reduce the modeling The number of Gaussian distributions is used to reduce the learning rate; when the brightness changes greatly, the number of Gaussian distributions is increased to increase the learning rate.

Preferably, the distribution parameters are matched and detected one by one with the current pixel value according to the priority, that is, to judge whether | μ _{i, t} -x _t |<max ⁽ Wσ _{i, t} , τ), where i=1, 2, ..., K, K is the number of Gaussian distributions of each pixel, μ _{i, t} and σ _{i, t} are the mean value and standard deviation of the i-th Gaussian distribution at time t, x _t is the previous pixel value, W and τ mean is the threshold constant.

Preferably, the target area detected in the previous frame is expanded and used as the target area of the current frame for matching detection, and the pixels outside the target area adopt a tight matching criterion, that is, both τ and W take larger values; The pixel points in the target area adopt a loose matching criterion, that is, both τ and W take smaller values, among which, 0.5<=W<=3.5, 3<=τ<=20.

Preferably, expand the target area of the previous frame by 10% as the target area of the current frame, take W=2.5, τ=15 for pixels outside the target area, and W=1.5, τ=15 for pixels inside the target area 6.

Preferably, when initializing at least one background model, the pixel values of each point in the first frame are used to initialize the mean value μ _K,0 of the Gaussian distribution, and the standard deviation σ _K,0 of the pixel values of each point in the first frame is set to 15<=σ _{K , 0} <= 25, the weight is 1/Kmax, and Kmax is the maximum number of Gaussian distributions for each pixel.

The method adopted in the present invention constructs a structured random measurement matrix by sampling and compressing images, reduces the amount of calculation data for Gaussian statistical modeling, and optimizes the efficiency of the algorithm in two aspects. One is to adaptively adjust the number of Gaussian models and the learning rate according to the change of the background brightness, and reduce the average calculation time; the other is to use different measurement values according to the segmentation and extraction of the target area of interest, and reduce the number of pixels involved in the modeling as a whole, effectively reducing the time for background modeling. The experimental results of algorithm simulation and node measurement prove that this method can obtain better target detection results and has strong anti-interference performance. Compared with the traditional mixed Gaussian algorithm, the memory capacity is reduced by about three-quarters, and the processing time can be reduced Reduced by more than 50%.

Description of drawings

What Fig. 1 shows is flow chart of the present invention;

What Fig. 2 shows is the comparison schematic diagram of the average processing time per frame of the present invention and different modeling methods;

What Fig. 3 shows is the comparison schematic diagram of false detection rate of the present invention and different modeling methods;

What Fig. 4 shows is the comparison schematic diagram of the missed detection rate of the present invention and different modeling methods;

What Fig. 5 shows is the performance of the present invention and different modeling methods and the comparison schematic diagram of average processing time per frame;

Fig. 6 is a schematic diagram showing the comparison of processing time and memory capacity between the present invention and different modeling methods.

detailed description

The technical solutions in the embodiments of the present invention will be clearly and completely described and discussed below in conjunction with the accompanying drawings of the present invention. Obviously, what is described here is only a part of the examples of the present invention, not all examples. Based on the present invention All other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

In order to facilitate the understanding of the embodiments of the present invention, specific embodiments will be taken as examples for further explanation below in conjunction with the accompanying drawings, and each embodiment does not constitute a limitation to the embodiments of the present invention.

For complex dynamic backgrounds, a linear combination of multiple Gaussian distributions can be used to accurately model the background. The mixed Gaussian model expresses the probability distribution characteristics of each pixel in the image as K (generally 3-5) Gaussian models. Each Gaussian distribution has different weights ωi, t (∑ωi, t=1, i=1, 2, . . . , K) and priorities (ω/σ), and are sorted from high to low in priority. When determining the background distribution, a certain threshold T (T<1) is taken for a single distribution. When the weight of the distribution is greater than or equal to the threshold, the Gaussian model is considered to be the background distribution, otherwise the Gaussian model is considered to be the foreground distribution. .

Let xt be a certain pixel value at time t, and its probability density function can be described by a linear combination of K Gaussian distributions:

Among them, ωi, t, μi, t and ∑i, t are the weight, mean and covariance gap matrix of the i-th Gaussian distribution at time t, respectively. Assuming that the color components of each pixel are independent of each other, the covariance matrix can be expressed as:

The K Gaussian distributions are arranged in descending order of ω/σ, and the Gaussian distribution whose weight is greater than a certain threshold represents the background distribution, namely:

f _i (x|μ, σ ² )∈Bg, if ω _i ＞TH (3)

Among them, TH is the background threshold.

1. Background initialization method

In general, the image sequence collected by the video node does not change much in a period of time, so it can be considered that the gray value of each pixel obeys the Gaussian distribution of mean value μ and standard deviation σ, and each image The Gaussian distributions of prime points are independent. Firstly, the background is initialized. In order to reduce the computational complexity and storage capacity, three Gaussian models are used to model each pixel (K=3). The Gaussian distribution expectation μ is initialized with the pixel values of each point in the first frame of the image, the standard deviation σ takes a larger value (σ=20), and the weight is initially set to 1/(2i+1) (i=0,1 ,2).

2. Background model learning and updating method

When performing target detection, match xt with each Gaussian distribution one by one according to the priority order ω/σ from large to small. If no Gaussian distribution representing the background distribution is detected that matches xt, the point is considered to be the target otherwise it is the background. The specific execution steps of the background model algorithm are as follows:

(1) Matching criteria

Compare K Gaussian distributions with the current pixel value xt according to priority to see if the condition |μ _i,t -x _t |<max(W*σ,λ), (i=1,2,...,K) , where W and λ are coefficients. Different conditions are used for the pixels in the target area and non-target area to determine the target and background. The specific operation is to enlarge the size of the target frame detected in the previous frame image by 1.1 times as the target frame of the current frame.

The following conditions are used to judge the pixels belonging to the target area:

|x- _μi |<max(1.5*σ, 6) (4)

The judgment conditions for the pixels in the non-target area are:

|x- _μi |<max(2.5*σ, 15) (5)

In addition, different value strategies are adopted according to the position of the pixel: for pixels outside the target area, W=2.5, λ=15; for pixels inside the target area, W=1.5, λ=6. In this way, the pixels in the object area are more likely to be detected as objects, increasing the shape integrity of object detection.

(2) Background learning and updating

Use the obtained pixel value to match with the known i-th Gaussian model, if the matching is successful, then update the matching i-th Gaussian model distribution parameters according to formula (6):

In equation (2.16) Among them, β is used to control the update speed of background and foreground (different values are taken according to whether the model describes the background). Foreground updates will be slower than background in most cases. The weights of the K Gaussian distributions are updated as follows:

ω _i,t+1 = (1-α _i )ω _i,t +α _i M _i,t (7)

In equation (7) The size of α determines its priority in the background and determines the update speed of the weights of each Gaussian component. The smaller α is, the more stable the background image is; the size of β determines the update speed of the background, and the larger the β, the faster the convergence speed of the background image.

The adaptive Gaussian mixture model (AGMM) algorithm steps are as follows:

1) Initialize the Gaussian distribution (weight, expectation, variance) with the first frame image, k=0;

2) For the new pixel at time t

Judge whether to match according to formula (4) (5);

If yes, go to step 3), otherwise go to step 4);

3) For the matched Gaussian model, use formula (6) (7) to update;

4) If it does not match, initialize a new Gaussian model with the current value (small weight, large variance);

5) Calculate the weight variance ratio ω/σ, arrange in descending order, and replace the minimum value;

6) According to |X _{t+1, i} -B _{i, t} |<T, judge whether the pixel is foreground or background, and output the result;

Go to step 2);

7) End, the next frame of image.

3. Post-processing method

According to the method described above, the binary image template Morg of the target can be obtained. Perform 3×3 morphological opening operation on the binary template Morg to obtain the result Ms, and then remove isolated points through 3×3 corrosion operation to obtain the result M. This process leads to the loss of some target pixels, and the following processing method based on morphological target reconstruction can retain as many target images as possible:

F in equation (8) is the final result after foreground extraction and noise filtering. The size of the structural element SE in the equation depends on the size of the detected object. Experiments show that the use of 3×3 structural elements can achieve better target detection results. Filling holes in the segmented foreground object F by using structural elements combined with assimilative filling can make the object more complete. Finally, the small blocks smaller than 40 pixels are removed through the statistical results of the target size to achieve the purpose of eliminating noise.

The Gaussian mixture model parameters are continuously updated to accommodate gradual changes in the background. In addition, because the algorithm performs 3 to 5 mixed Gaussian modeling for each pixel in the image, the overall calculation and storage capacity are relatively large.

4. Structured compressed sensing algorithm

At present, the commonly used modeling method is to describe the dynamic background by using the mixed Gaussian background model. Since 3 to 5 Gaussian models are established for each pixel, the method of mixing Gaussian models consumes a large amount of computing and storage resources of the camera node, which affects the real-time application of the algorithm. In order to improve the efficiency of the algorithm, the compressed sensing algorithm is introduced to randomly sample the image data, thereby reducing the amount of calculation and storage of the background modeling algorithm. However, the completely random nature of the random sampling matrix makes the implementation of the hardware circuit more complex and the target detection results are uncertain. In view of this situation, the present invention introduces the structured compressed sensing algorithm into the mixed Gaussian modeling, and on this basis, studies an adaptive mixed Gaussian background modeling method that uses a structured random measurement matrix to sample images, and Optimize the overall performance of the algorithm to improve the overall operating efficiency.

Compressed sensing uses a measurement matrix Φ of M rows and N columns (M<<N) to measure the signal x (N dimension), and the compressed measurement value y (M dimension) can be obtained. This process can be expressed by equation (9) accomplish.

y=φx=φΨα=Θα (9)

If the signal x has sparsity in a certain variation domain, as shown in equation (10):

α＝ ^ΨT x (10)

And the measurement matrix Φ satisfies the constrained isometric condition, that is, for any K-sparse signal f and the constant δ _k ∈ (0, 1) satisfies: Then the signal can be perfectly restored by equation (11):

This process is called reconstruction, where the 0 norm refers to the number of 0 elements.

There are three types of measurement matrices proposed so far that satisfy the condition of constrained equidistantness. The first category includes Gaussian random measurement matrices, Bernoulli random matrices, etc. whose matrix elements independently obey a certain distribution. The second category includes partially orthogonal matrices, partially Hadamard matrices, and uncorrelated measurement matrices. Such matrices are only uncorrelated with signals that are sparse in the time or frequency domain. The third category includes Toeplitz matrix, structured random measurement matrix, Chirps measurement matrix, circular matrix, and perceptual matrix formed by random convolution.

1) Random Gaussian matrix: As shown in formula (12), each element of the measurement matrix independently obeys a Gaussian distribution with a mean value of 0 and a variance of 1/M, and the value is 1 or 0 with equal probability. The advantage of a Gaussian measurement matrix is that the number of measurement rows required is small and it is almost uncorrelated to arbitrarily sparse signals.

2) Random Bernoulli matrix; as shown in formula (13), each element of the measurement matrix independently obeys a symmetric Bernoulli distribution, and the value is 1 or -1 with equal probability. This matrix is very random and has properties similar to Gaussian matrices.

3) Partially orthogonal matrix; the step of constructing the matrix is to first generate an N×N orthogonal matrix U, then randomly select M row vectors in the matrix U and unitize the column vectors of the M×N matrix, then Get a partially orthogonal matrix.

4) Toeplitz matrix; the step of constructing the matrix is to first generate the measurement matrix Φ, in each row vector of the matrix Φ, randomly select the position according to the probability distribution of the elements of equation (14), and then assign 0 to the corresponding position , 1, -1, where

5) Partial Hadamard matrix; the step of constructing this matrix is to first generate a Hadamard matrix with a size of N×N, and then randomly select M row vectors in the generated matrix to form an M×N Hadamard measurement matrix.

6) Structured random measurement matrix; Although random Gaussian and random Bernoulli matrices are non-correlated to many sparse signals, due to their completely random characteristics, calculations are more complicated and memory capacity requirements are high. Therefore, many studies have proposed the concept of structured random measurement matrix. The construction of this type of matrix adopts the mixed model of random Gaussian, Bernoulli matrix and partial Fourier transform matrix, randomly selects M rows from the N×N mixed matrix, and then normalizes each column. The structured random measurement matrix is almost independent of all other orthogonal matrices and maintains the advantages of various matrices.

7) Deterministic matrix; completely random matrix has the disadvantages of uncertain factors and difficult implementation of hardware circuits. In order to overcome its shortcomings in compressed sensing applications, many studies have proposed deterministic measurement matrices, including polynomial deterministic matrices and rotation measurement matrices. Wait.

The present invention comprises steps:

Image reconstruction step: divide the image into blocks according to the size of the collected images, and convert the collected image blocks into N×1 vectors;

Compressive sensing step: construct a structured random measurement matrix to sample and compress the converted vector;

Background modeling step: use the adaptive mixed Gaussian model to perform Gaussian modeling on each measured matrix block, and use the least pixel rule to detect the target block and background block; use the mixed Gaussian model to establish at least one pixel for each pixel of each candidate node. Background model, initialize the background model with the first frame of image data, set a uniform background threshold for each background model, the background model whose pixel weight is greater than the background threshold describes the background distribution, and the pixel weight is less than or equal to The background model of the background threshold describes the foreground distribution. Use the pixels judged as the background in the image to reinitialize the background model whose weight is less than the initialization threshold; Matching detection, determine that the pixel points whose background models do not match the current pixel values are points in the target area, update the distribution parameters for the successfully matched background models, and update the weights for each background model;

Update step: use different strategies to update the target block and background block, and adjust the parameters of the structured random measurement matrix according to the detection results;

Post-processing step: performing post-processing on the detected target image to obtain the final target image;

Among them, different strategies include setting different sampling values M according to the region of interest, increasing the sampling rate in the area of 1.2 times the target block in the previous frame, and reducing the sampling rate in the background area; and, when the background brightness changes little, reduce the modeling The number of Gaussian distributions is used to reduce the learning rate; when the brightness changes greatly, the number of Gaussian distributions is increased to increase the learning rate.

The distribution parameters are matched and detected one by one with the current pixel value according to the priority, that is, to judge whether |μ _{i, t} -x _t |<max(Wσ _{i, t} , τ), where i=1, 2, ..., K, K is the number of Gaussian distributions of each pixel, μ _{i, t} and σ _{i, t} are the mean and standard deviation of the i-th Gaussian distribution at time t, respectively, x _t is the previous pixel value, W and τ are threshold constants.

The target area detected in the previous frame is expanded and used as the target area of the current frame for matching detection. The pixels outside the target area adopt a tight matching criterion, that is, both τ and W take larger values; The pixel points adopt a loose matching criterion, that is, both τ and W take smaller values, among which, 0.5<=W<=3.5, 3<=τ<=20. Expand the target area of the previous frame by 10% as the target area of the current frame, set W=2.5, τ=15 for pixels outside the target area, and W=1.5, τ=6 for pixels inside the target area.

When initializing at least one background model, the pixel value of each point in the first frame is used to initialize the mean value of the Gaussian distribution μ _{K, 0} , and the standard deviation σ _{K, 0} of the pixel value of each point in the first frame is 15<=σ _{K, 0} <= 25, the weight is 1/Kmax, and Kmax is the maximum number of Gaussian distributions for each pixel.

The object detection algorithm of the present invention first divides the image xt collected by the video node into 4×4 or 8×8 blocks, and then constructs a structured random measurement matrix Φ to directly sample the image in the space domain to obtain a compressed image yt. According to the compressed sensing theory, yt contains most of the information of the original image. The background model is constructed through the adaptive Gaussian mixture model (AGMM), and the foreground image is obtained through background subtraction, and then the foreground image is morphologically processed.

This application first selects the active node for target detection and target tracking, and then selects the current optimal node for target tracking through the performance function f(i), as shown in Figure 1. The target detection adopts adaptive Gaussian mixture background construction Realize the detection and segmentation of moving targets; implement target tracking and state estimation of nodes through distributed mean shift and target association, combine the detection effect of sensor nodes, communication energy consumption and other factors to determine the sensor network performance evaluation function, and select the optimal sensor Nodes for object tracking. Comprehensive consideration of computational complexity, data transmission, and storage requirements enables accurate tracking of moving targets in complex scenes in a wide range. As shown in Figure 2 to Figure 6, according to comparison with other existing algorithms, the present invention can obtain better target detection results and has stronger anti-interference. Compared with the traditional mixed Gaussian algorithm, the memory capacity is reduced by about four Three-thirds, the processing time can be reduced by more than 50%.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present invention can easily think of changes or Replacement should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims (5)

1. a kind of object detection method towards video-aware node, it is characterised in that including step：

Image reconstruction step：Picture size size according to collecting carries out piecemeal to image, will adopt the image block conversion for obtaining For the vector of N × 1；

Compressed sensing step：Build structuring random measurement matrix carries out Sampling Compression to the vector after conversion；

Background modeling step：Gauss modeling is carried out to the matrix-block after each measurement using ADAPTIVE MIXED Gauss model, is adopted Object block and background block detection are carried out with minimum pixel rule；Each pixel of each both candidate nodes is set up by mixed Gauss model At least one background model, is initialized with the first frame image data to the background model, to each background model setting Unified background threshold, pixel weights are background distributions, pixel weights more than the background model description of the background threshold It is distributed for prospect less than or equal to what the background model of the background threshold was described, it is again initial with the pixel for being judged as background in image Change background model of the weights less than initial threshold value；The distributed constant of background model is according to priority current with corresponding from big to small Pixel value matching detection one by one, judge background model with the unmatched pixel of current pixel value as the point in target area, Distributed constant is updated to the background model that the match is successful, weight is updated to each background model；

Update step：Object block and background block are updated using different strategies, and according to the result for detecting to structuring Random measurement matrix carries out parameter regulation；

Post-processing step：Target image to detecting carries out later stage process and obtains final target image；

Wherein, different strategies includes arranging different sampled values M according to interest region, in 1.2 times of areas of object block of former frame Sample rate is improved in domain, and reduces sample rate in background area；And, when background luminance changes less, the Gauss point for reducing modeling Cloth number, to reduce learning rate；When brightness is changed greatly, Gaussian Profile number is improved, to improve learning rate.

2. the object detection method towards video-aware node according to claim 1, it is characterised in that distributed constant is pressed Priority carries out matching detection one by one with current pixel value, that is, discriminate whether to meet | μ_i,t-x_t| ＜ max (W σ_i,t, τ), i in formula =1,2 ..., K, K are the number of each pixel Gaussian Profile, μ_i,tAnd σ_i,tRespectively i-th Gaussian Profile of t average and Standard variance, x_tFor preceding pixel value, W and τ is threshold value constant.

3. the object detection method towards video-aware node according to claim 2, it is characterised in that examine previous frame The target area measured carries out matching detection after extension as the target area of present frame, the pixel outside target area Higher value is taken using tight matching criterior, i.e. τ and W；Pixel in target area is using loose matching criterior, i.e. τ and W Smaller value is taken, wherein, 0.5<=W<=3.5,3<=τ<=20.

4. the object detection method towards video-aware node according to claim 3, it is characterised in that by previous frame mesh Region extension 10% is marked as the target area of present frame, the pixel outside target area takes W=2.5, τ=15, in target Pixel in region takes W=1.5, τ=6.

5. the object detection method towards video-aware node according to claim 1, it is characterised in that initialization is at least During one background model, it is used for initializing Gaussian Profile mean μ by the first frame each point pixel value_K,0, the first frame each point pixel value Standard variance σ_K,0Take 15<=σ_K,0<=25, weight is 1/Kmax, and Kmax is the largest Gaussian one distribution number of each pixel.