CN106023258B - An Improved Adaptive Gaussian Mixture Model Moving Object Detection Method - Google Patents

Abstract

The present invention provides a kind of improved adaptive GMM moving target detecting method, Gaussian Profile and parameter is initialized, to k Gaussian Profile according to weight ω_i,tSize be ranked up, wherein [1, K] i ∈；Update distribution, according to whether match with existing Gaussian Profile, be updated, including more new peak weight distribution, merge weight close to be distributed, delete low weight distribution, increase new distribution；The value of this kind of improved adaptive GMM moving target detecting method, learning rate and model profile number is updated, and solves the problems, such as data redundancy；Meanwhile parameter value is updated, it can preferably adapt to environmental change.This method is the moving object detection algorithm based on Video Image processing technique, can be applied to Real-time Video Processing System, establishes critical theoretical and application foundation for vehicle detection or fire monitoring, has extraordinary application prospect.

Description

An Improved Adaptive Gaussian Mixture Model Moving Object Detection Method

technical field

The invention relates to an improved adaptive Gaussian mixture model moving target detection method.

Background technique

At present, optical flow method, frame difference method, and background subtraction method are commonly used algorithms for moving target detection. Among them, the optical flow method is computationally complex, requires special hardware support, and has poor real-time performance and practicability. Although the frame difference method can effectively remove the static background, the extracted target is often rough and larger than the actual moving target contour, and there will be holes and "double shadows" in the target. The background subtraction method can detect moving objects when the environment changes, but it needs to update the background image in real time.

From a practical application point of view, background subtraction is the most widely used method for moving object detection. Among them, the most commonly used background model is the Gaussian mixture model. However, the traditional Gaussian mixture model, due to its relatively fixed learning rate and model distribution number, is prone to data redundancy; at the same time, the relatively fixed parameter values still have problems such as not being able to better adapt to environmental changes.

The above problems are problems that should be considered and solved in the process of moving target detection.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an improved adaptive Gaussian mixture model moving target detection method to solve the problem in the prior art that due to the relatively fixed values of the learning rate and the model distribution number, it is easy to cause data redundancy; There are also problems such as not being able to better adapt to environmental changes in the value of the parameters.

The technical solution of the present invention is:

An improved adaptive Gaussian mixture model moving target detection method, comprising:

S1. Initialize the Gaussian distribution and parameters, and sort the k Gaussian distributions according to the size of the weight ω _{i, t} , where i ∈ [1, K];

S2, update the distribution, according to whether it matches the existing Gaussian distribution, update, including updating high-weight distribution, merging weight close distribution, deleting low-weight distribution, adding new distribution, specifically:

S21. If the pixel value does not match the existing distribution, if the K value of the distribution number does not reach the maximum value, add a Gaussian distribution; if the K value of the distribution number has reached the maximum value, delete the Gaussian distribution with the smallest weight, and then add another Gaussian distribution. Add a Gaussian distribution;

If there is a Gaussian distribution that satisfies the formula (10), that is, the mean value is similar, the distribution is merged, the distribution with a larger weight is updated, and the distribution with a smaller weight is deleted;

|μ _x -μ _y |≤T _μ (10)

Among them, T _μ is the threshold of the minimum interval of the mean, and μ _x and μ _y represent the mean of two different distributions;

S22. After processing the distribution number and the mean, the weight is judged; if the weight is lower than the threshold, the Gaussian distribution that does not meet the current requirements is discarded.

Further, in step S1, the initialization of the parameter mean value μ ₀ is specifically:

Parametric variance The initialization is specifically:

In formulas (1) and (2), N is the number of video images; X _t is the value of the pixel at time t.

Further, in step S1, the initialization of the distribution number K ₀ is specifically:

Among them, K _L , K _M , K _S are the Gaussian distribution numbers of three levels respectively, T _K1 , T _K2 are the thresholds required for assigning the initial distribution numbers; the value of K ₀ is 1-5.

Further, in step S21, the weight update calculation formula is as follows:

ω _k,t =(1-α _k,t )*ω _k,t-1 +α _k,t *M _k,t (5)

Among them, ω _k,t is the weight of the kth Gaussian mixture model at time t, α _k,t is the learning rate of the kth Gaussian mixture model at time t, M _k,t is the matching factor, for the matching model M _{k ,t} =1; otherwise, M _k,t =0.

Further, in step S21, for the unmatched model, the mean μ and the variance σ ² remain unchanged; for the matched model, the parameter update formula is as follows:

ρ _k,t =α _k,t *η(X _t , μ _k,t ,σ _k,t ) (6)

μ _k,t =(1-ρ _k,t )·μ _k,t-1 +ρ _k,t X _t (7)

Among them, ρ _k,t is the parameter update factor of the kth Gaussian mixture model at time t; η(X _t , μ _k,t ,σ _k,t ) is the Gaussian mixture model at time t, and its calculation formula is as follows:

in, δ _{i, t} is the variance, I is a two-dimensional unit matrix, X _t is the pixel value of the point at time t, μ _{i, t} is the mean of the ith Gaussian mixture model at time t.

Further, the learning rate α _k,t is updated as follows:

Among them, α _{k, t} is the value of the k-th learning rate at time t; α ₀ is the initial learning rate, ranging from 0.01 to 0.1, and f _{k, t} is the number of changes of the Gaussian distribution per unit time, that is, the update frequency ; T _f is the threshold for judging the update frequency.

The beneficial effects of the invention are as follows: in the improved adaptive Gaussian mixture model moving target detection method, the values of the learning rate and the model distribution number are updated to solve the problem of data redundancy; better adapt to environmental changes. The method is a moving target detection algorithm based on computer video image processing technology, which can be applied to real-time video processing systems, lays a key theoretical and application foundation for vehicle detection or fire monitoring, and has very good application prospects.

Description of drawings

FIG. 1 is a schematic flowchart of an adaptive Gaussian mixture model moving object detection method improved by an embodiment.

FIG. 2 is a schematic flowchart of distribution update in the adaptive Gaussian mixture model moving object detection method improved by the embodiment.

Figure 3 is a source image acquired in an embodiment.

Figure 4 is an image of the detection result of the traditional Gaussian mixture model.

FIG. 5 is an image of the detection result of the embodiment method.

Detailed ways

The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Example

The embodiment comprehensively uses a variety of video image processing methods, and designs an improved adaptive Gaussian mixture model moving target detection method by analyzing the traditional Gaussian mixture model in detail.

The characteristics of moving objects in video images include the following: the moving objects have obvious differences with the color of the background objects; the moving speed or frequency of the moving objects is within the range that can be observed by the naked eye; the changes in external conditions such as lighting in the video are relatively gentle. The improved adaptive Gaussian mixture model moving object detection method proposed in the embodiment is developed according to the above-mentioned characteristics of moving objects in video images. The details are as follows:

Initialization of mean and variance

The calculation formula of the initial mean μ ₀ is as follows:

initial variance The calculation formula is as follows:

Among them, N is the number of video images; X _t is the value of the pixel at time t.

distribution number initialization

The formula for calculating the distribution number K ₀ is as follows:

Among them, _KL , _KM , and KS are the three-level Gaussian distribution numbers, respectively, and T _K1 and _{T K2} are the thresholds required to assign the initial distribution numbers; the value of K ₀ is generally 1 to 5.

Learning rate update

The learning rate α _{k, t} is updated as follows:

Among them, α _{k, t} is the value of the k-th learning rate at time t; α ₀ is the initial learning rate, ranging from 0.01 to 0.1, and f _{k, t} is the number of changes of the Gaussian distribution per unit time, that is, the update frequency ; T _f is the threshold for judging the update frequency.

Among them, α _{i, t} is the value of the i-th learning rate at time t; α ₀ is the initial learning rate, which is generally 0.01 to 0.1, and f is the number of changes in the Gaussian distribution per unit time, that is, the update frequency; T _f is the threshold for judging the update frequency.

Weight, mean, variance update

The weight update calculation formula is as follows:

ω _k,t =(1-α _k,t )*ω _k,t-1 +α _k,t *M _k,t (5)

Among them, ω _k,t is the weight of the kth Gaussian mixture model at time t. M _k,t is the matching factor, M _k,t =1 for a matched model; otherwise, M _k,t =0.

For unmatched models, the mean μ and variance σ ² are unchanged.

For the matched model, its parameter update formula is as follows:

ρ _k,t =α _k,t *η(X _t , μ _k,t ,σ _k,t ) (6)

μ _k,t =(1-ρ _k,t )·μ _k,t-1 +ρ _k,t X _t (7)

Among them, ρ _k,t ρ is the parameter update factor; η(X _t , μ _k,t ,σ _k,t ) is the Gaussian mixture model at time t, and its calculation formula is as follows:

in, δ _i,t is the variance, and I is the two-dimensional identity matrix.

Gaussian model update

The establishment process of the improved Gaussian mixture model is shown in Figure 1 below. The k Gaussian distributions are sorted according to the size of the weight ω _{i, t} , where i ∈ [1, K]. According to whether it matches the existing Gaussian distribution, update it according to Figure 2. If the pixel value does not match the existing distribution, if the distribution number K value does not reach the maximum value, a new Gaussian distribution will be added; if the distribution number K value has reached the maximum value value, delete the Gaussian distribution with the smallest weight, and add another Gaussian distribution.

If there is a Gaussian distribution that satisfies the formula (10), that is, the mean values are similar, the distributions are merged, the distributions with larger weights are updated, and the smaller ones are deleted.

|μ _x -μ _y |≤T _μ (10)

Among them, T _μ is the threshold of the minimum interval of the mean.

After processing the distribution number and the mean, the weight is judged. If the weight is lower than the threshold, the Gaussian distribution that does not meet the current requirements will be discarded.

Experimental verification

The embodiment method is based on the Microsoft Visual Studio 2010 platform, and uses the Opencv computer vision library to develop, import the original video image Figure 3, and use the Gaussian mixture model that comes with Opencv to detect moving objects in Figure 3 to obtain Figure 4, using the embodiment method Figure 5 is obtained by detecting moving objects in Figure 3. It can be seen that Figure 5 has obvious edge detail information compared to Figure 4.

Claims (4)

1. an improved adaptive Gaussian mixture model moving target detection method, is characterized in that, comprises: S1. Initialize the Gaussian distribution and parameters, and sort the k Gaussian distributions according to the weights ω _{i, t} , where i∈[1, K], ω _{i, t} is the weight of the ith Gaussian mixture model at time t ; S2, update the distribution, according to whether it matches the existing Gaussian distribution, update, including updating the weight of the Gaussian distribution, merging the weights close to the distribution, deleting the low-weight distribution, and adding a new distribution, specifically: S21. If the pixel value does not match the existing distribution, if the K value of the distribution number does not reach the maximum value, add a Gaussian distribution; if the K value of the distribution number has reached the maximum value, delete the Gaussian distribution with the smallest weight, and then add another Gaussian distribution. Add a Gaussian distribution; If there is a Gaussian distribution that satisfies the formula (10), that is, the mean value is similar, the distribution is merged, the distribution with a larger weight is updated, and the distribution with a smaller weight is deleted; |μ _x -μ _y |≤T _μ (10) Among them, T _μ is the threshold of the minimum interval of the mean, and μ _x and μ _y represent the mean of two different distributions; In step S21, for the unmatched model, the mean μ and the variance σ ² remain unchanged; for the matched model, the parameter update formula is as follows: ρ _k,t =α _k,t *η(X _t , μ _k,t ,σ _k,t ) (6) μ _k,t =(1-ρ _k,t )·μ _k,t-1 +ρ _k,t X _t (7) Among them, ρ _k,t is the parameter update factor of the kth Gaussian mixture model at time t; α _k,t is the learning rate of the kth Gaussian mixture model at time t; η(X _t , μ _k,t ,σ _{k, t} ) is the Gaussian mixture model at time t, and its calculation formula is as follows: in, δ _{i, t} is the variance, I is a two-dimensional unit matrix, X _t is the value of the pixel at time t, μi, t is the mean of the ith Gaussian mixture model at time t; The learning rate α _{k, t} is updated as follows: Among them, α _{k, t} is the value of the k-th learning rate at time t; α ₀ is the initial learning rate, ranging from 0.01 to 0.1, and f _{k, t} is the number of changes of the Gaussian distribution per unit time, that is, the update frequency ; T _f is the threshold for judging the update frequency; S22. After processing the distribution number and the mean value, the weight value is judged instead of updating; if the weight value is lower than the threshold value, the Gaussian distribution that does not meet the current requirements is discarded. 2. improved adaptive Gaussian mixture model moving target detection method as claimed in claim 1, is characterized in that, in step S1, the initialization of parameter mean value μ ₀ is specifically: Parametric variance The initialization is specifically: In formulas (1) and (2), N is the number of video images; X _t is the value of the pixel at time t. 3. improved adaptive Gaussian mixture model moving target detection method as claimed in claim 1, is characterized in that, in step S1, the initialization of distribution number K ₀ is specifically: Among them, K _L , K _M , K _S are the Gaussian distribution numbers of three levels, respectively, is the parameter variance, T _K1 and T _K2 are the thresholds required to assign the initial distribution number; the value of K ₀ is 1-5. 4. the improved adaptive Gaussian mixture model moving target detection method as claimed in any one of claims 1-3, is characterized in that, in step S21, the update calculation formula that the distribution that the weight value is larger is carried out update processing is as follows: ω _k,t =(1-α _k,t )*ω _k,t-1 +α _k,t *M _k,t (5) Among them, ω _{k, t} is the weight of the k-th Gaussian mixture model at time t, α _{k, t} is the learning rate of the k-th Gaussian mixture model at time t, M _{k, t} is the matching factor, for the matching model M _{k ,t} =1; otherwise, M _k,t =0.