Real-time detection method of high-altitude projectiles based on improved Gaussian mixture algorithm

Abstract

The present invention relates to a real-time detection method for high-altitude objects based on an improved Gaussian mixture algorithm, and a high-altitude object detection and positioning coordinate based on an improved Gaussian mixture model. The image of a building in a video frame shot by a camera is processed into a grayscale image, and the pixel points are subjected to Gaussian model statistics. The next frame of the image is intercepted, and the difference of the Gaussian function of each pixel point with respect to the background model is analyzed. If the match is successful, the pixel point is set to 0, otherwise it is considered to be a foreground pixel point. The background pixel statistical model is updated so that each frame of the image can be background-fused. The next frame of the image is continued, and finally all the coordinates of the moving object in the video are obtained. The algorithm solves the interference of diversified dynamic scenes and can effectively reduce the false alarm rate of high-altitude object detection. The method can well detect and track moving targets, and has good anti-interference ability and good accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of video detection, and relates to a real-time detection method for high-altitude objects based on an improved Gaussian mixture algorithm, and is a high-altitude objects monitoring and positioning method based on a mixed Gaussian model that uses an improved mixed Gaussian model to reduce the false alarm rate of high-altitude objects detection. Background Art

[0002] With the improvement of people's quality of life and the progress of social technology, more and more intelligent technologies are pouring into all aspects of community life, especially with the current progress of urbanization, which shows a dense population and large-scale. However, with the advancement of urbanization, some problems have also been highlighted, such as high-altitude throwing, which has become a bomb hanging over the city. The serious consequences of some recent high-altitude throwing incidents have also attracted people's attention to high-altitude throwing, and more and more high-altitude throwing technologies and applications have emerged. For example, image video processing methods and hardware sensing equipment are used for high-altitude throwing warning and intelligent video evidence collection, and efforts are intensified to monitor and eliminate potential high-altitude throwing dangerous goods in buildings. In terms of image intelligent high-altitude throwing processing, during detection, due to many interference factors, diverse and complex scenes, unstable light and other objective factors, the false alarm rate of intelligent detection is very high.

[0003] The present invention proposes a moving target detection method using an improved Gaussian mixture model, which is different from the background subtraction method of a common moving target detection algorithm, improves the anti-interference ability, and strengthens the recognition of dynamically changing backgrounds. Summary of the invention

[0004] Technical issues to be solved

[0005] In order to avoid the shortcomings of the prior art, the present invention proposes a real-time detection method for high-altitude objects based on an improved Gaussian mixture algorithm to solve the problems in the prior art and provide a mobile target detection method based on an improved Gaussian mixture model that can eliminate other interference factors in the video and reduce the false alarm rate of high-altitude object detection; by intelligently solving the coordinates of objects thrown from high altitudes and dynamically updating the background model, the interference of diverse scenes such as trees, flying birds, and window glass can be eliminated, thereby effectively reducing the false alarm rate of high-altitude object detection.

[0006] Technical Solution

[0007] A real-time detection method for high-altitude projectiles based on an improved Gaussian mixture algorithm, characterized by the following steps:

[0008] Step 1: Use the Gaussian function of each pixel point of the previous N frames to establish a Gaussian mixture background statistical model:

[0009]

[0010] where μ k is the mean, D is a constant parameter, x is the pixel value of each image, is the covariance of the kth component, I is the RGB color space, η(x; θ k ) is the kth Gaussian normal distribution; k distributions are distributed according to the fitness value w k / σ k ,w k is the kth Gaussian component weight coefficient, σ k is the mean square error;

[0011] Step 2: The first frame image in the video after the first N frames, that is, the N+1th frame image, is gray-processed according to the adaptive threshold to convert the RGB image into a gray-scale image;

[0012] The algorithm is divided into three channels and runs in three parallel processes; the final partition F(xi) comes from the three parts F R (x i ),F B (x i ),F G (x i ) are operated bit by bit or one by one to obtain the required grayscale value;

[0013] Step 3: Compare the background statistical model in step 1 with the Gaussian mixture model of the N+1th frame image, and add a dynamic background to the N+1th frame image:

[0014] An adaptive learning rate is introduced to reflect the speed at which the foreground image merges into the background, and the estimate of t samples is recorded as Recursively write the estimation function For the t-1 sample and the ownership coefficient The last sample:

[0015] Use adaptive learning rate to classify frequently moving foreground as dynamic background;

[0016] Using Dynamic Estimator Make adjustments

[0017] Get the adaptive threshold;

[0018] Where T(x i ) is a dynamic learning controller, T inc ,T dec is a constant parameter;

[0019] Step 4: Extract each pixel in the N+1th frame again, and use the Bayesian decision rule and likelihood function

[0020]

[0021] Analyze the difference between the Gaussian mixture model parameters of the pixel and the background model, and set the threshold;

[0022] Among them: BG is the background, FG is the foreground, is the background or foreground probability at x pixels;

[0023] Step 5: Set the pixel within the threshold to 0, which is the background pixel; otherwise, set the pixel to 255, which is the foreground pixel;

[0024] Step 6: Update the Gaussian Mixture Model

[0025] Adopting the adaptive Gaussian mixture model, the Gaussian distribution parameters of the foreground pixels, namely the mean, standard deviation, and weight, are transformed into

[0026] Added to the background statistics model:

[0027]

[0028] And update according to the following update equation:

[0029]

[0030] in: is the estimated mean, is the estimated variance, used to describe the Gaussian component; is the mixing weight, and is non-negative and adds up to 1;

[0031] If a pixel only appears once, reduce its weight. When the weight is less than the threshold, delete it from the mixed Gaussian model.

[0032] Step 7: Continue to read the next frame of video, repeat steps 2 to 6, and finally obtain the foreground pixels of all frames of video; finally, return the selected target to the video to achieve real-time target tracking of high-altitude objects.

[0033] The specific steps of step 6 "updating the background mixed Gaussian model" are:

[0034] S6a introduces the adaptive model parameter learning rate α and sets the value of the learning rate;

[0035] S6b adopts a mechanism to improve the convergence rate of the mixed Gaussian model. For each Gaussian distribution, its historical matching number is introduced into the calculation of the model learning rate;

[0036] At the beginning of background modeling, S6c averages the learning rate of each frame image as the background, and when the total number of video frames is less than 1 / α, each frame updates each distribution in the Gaussian mixture model with the same time weight, avoiding the first frame image from having too high a weight in the mixed Gaussian background model;

[0037] In order to avoid the newly added Gaussian distribution being deleted as a low-probability event, which affects the normal background modeling and reduces the robustness of the system, S6d adopts a model for adaptively updating the Gaussian score. When deleting the Gaussian distribution, the average weight of the Gaussian distribution and the number of frames from the latest matching frame of a Gaussian distribution to the current frame are taken into account.

[0038] The value of the S6a learning rate is related to the speed at which the foreground image merges into the background. When the learning rate is large, the speed at which the foreground image merges into the background is greater than when the learning rate is small.

[0039] Beneficial Effects

[0040] The present invention proposes a real-time detection method for high-altitude objects based on an improved Gaussian mixture algorithm, and a high-altitude object detection and positioning coordinate based on an improved Gaussian mixture model. The image of the building in the video frame captured by the camera is processed into a grayscale image, and the pixel points are statistically analyzed by a Gaussian model. The difference of the Gaussian function of each pixel point with respect to the background model is intercepted, and if the match is successful, the pixel point is set to 0, otherwise it is considered to be a foreground pixel point. The background pixel statistical model is updated so that each frame of the image can be background fused. Continue with the next frame of the image, and finally obtain all the coordinates of the moving object in the video. The algorithm solves the interference of diversified dynamic scenes and can effectively reduce the false alarm rate of high-altitude object detection.

[0041] The present invention adopts several Gaussian models to represent the characteristics of each pixel in the moving object image. The updating principle of the Gaussian mixture model is that if each pixel in the image matches the Gaussian mixture model, it is regarded as a background point, otherwise it is a foreground point. By updating the foreground model and calculating the short-term motion stability index, the effect of moving target detection is improved. By determining the relationship between the Gaussian distribution and the pixel, a new parameter is set to construct the background model. The analysis results show that this method can detect and track moving targets well, has good anti-interference ability and good accuracy.

[0042] Effect of the invention: According to the problems existing in the actual background modeling of the mixed Gaussian model, the traditional mixed Gaussian model is improved from two aspects: adaptive adjustment of the model parameter learning rate and the number of Gaussian distributions. The algorithm is experimentally tested using actual surveillance video, and the accuracy rate reaches more than 95%. The improved mixed Gaussian model improves the robustness, accuracy and real-time performance of the background modeling algorithm, and provides an accurate foreground area for the subsequent tracking and identification of moving targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the moving target detection algorithm.

[0044] Figure 2 This is a picture showing the effects of objects thrown from high altitude.

[0045] Figure 3 is the grayscale detection image. DETAILED DESCRIPTION

[0046] The present invention will now be further described with reference to the embodiments and the accompanying drawings:

[0047] like Figure 1 The figure shows a flowchart of a mobile target detection algorithm positioning method based on an adaptive Gaussian mixture model of the present invention. The method includes the following steps:

[0048] S1 starts the moving target detection of high-altitude dropped objects video, obtains the first N video frames in the video, and performs grayscale processing to obtain a grayscale image.

[0049] First, the algorithm will capture the first N frames of images and build a background statistical model based on the Gaussian function of each pixel. where μ k is the mean, is the covariance of the kth component, x is the pixel value of each image, η(x; θ k ) is the kth Gaussian normal distribution. The k distributions are distributed according to the fitness value w k / σ k ;

[0050] After S2 completes the acquisition of the first N frames of the video, it uses K Gaussian models to build a statistical model for the features of each pixel in the moving target image, and introduces an adaptive learning rate to reflect the speed at which the foreground image merges into the background.

[0051] Start the moving target detection of high-altitude parabolic video, obtain the N+1th frame image in the video, perform grayscale processing according to the adaptive threshold, and convert the RGB image into a grayscale image. Run the three channel distribution and run the algorithm in three parallel processes. The final segmentation F(xi) comes from the three parts F R (x i ),F B(x i ),F G (x i ) one by one by bitwise OR operation to get the required grayscale value.

[0052] S3 obtains the next frame of video, extracts each pixel in the video, and analyzes the difference between the Gaussian mixture model parameters of the pixel and the background model.

[0053] After the N+1th frame of the video is acquired, the background statistical model obtained in the first step is compared with the Gaussian mixture model of the frame to add a dynamic background: an adaptive learning rate is introduced to reflect the speed at which the foreground image merges into the background, and the estimate of t samples is recorded as It can be recursively written as the estimation function For the t-1 sample and the ownership coefficient The last sample: And use adaptive learning rate to classify frequently moving foreground (such as swaying trees) as dynamic background. Use dynamic estimator Make adjustments;

[0054]

[0055] Thus, the effect of adaptive threshold is achieved. i ) is a dynamic learning controller, T inc ,T dec It is a constant parameter. It has different values ​​in different situations, because in most cases the detected image is the background.

[0056] S4 considers the successfully matched pixel point as a background pixel point and sets the pixel point to 0, otherwise it sets the pixel point to 255, such as Figure 3 The white object shown is the detected moving target, and the rest is the background. Use OpenCV library functions to identify the grayscale image and save the coordinates in a list. Specifically:

[0057] S5 extracts each pixel in the N+1th frame again, and uses the Bayesian decision rule and likelihood function.

[0058]

[0059] Analyze the difference between the Gaussian mixture model parameters of the pixel and the background model, and set the difference threshold. BG is the background, FG is the foreground, is the background or foreground probability at x pixels.

[0060] S6 needs to update the background pixel statistical model at this time, using an adaptive Gaussian mixture model. For pixels that fail to match, the Gaussian distribution parameters (mean, standard deviation, weight) of the current pixel are added to the mixed Gaussian model. If a pixel only appears once, its weight is reduced. When the weight is less than the threshold, it is deleted from the mixed Gaussian model. Specifically:

[0061] At this time, the background statistical model needs to be updated. An adaptive Gaussian mixture model is used. For the foreground pixel, the Gaussian distribution parameters (mean, standard deviation, weight) of the current pixel are selected in a suitable time period T, and at time t, there is Added to the background statistical model

[0062]

[0063] And update according to the following update equation

[0064] in is the estimated mean, is the estimated variance, describing the Gaussian component. is the mixed weight, and is non-negative and adds up to 1. If a pixel appears only once, reduce its weight, and when the weight is less than the threshold, delete it from the mixed Gaussian model.

[0065] S7 continues to read the next frame of video. Finally, the coordinates of all foreground points are obtained, and the original video is output to select the moving target.

[0066] According to the mobile target detection algorithm based on the adaptive Gaussian mixture model, it is characterized in that: the specific steps of step S6 "updating the background mixture Gaussian model" are:

[0067] S6a introduces an adaptive model parameter learning rate α, and sets a value for the learning rate. If the value is large, the foreground image will blend into the background too quickly, resulting in a hole or incompleteness in the moving target area; on the contrary, if the value is small, the newly appearing interference motion in the complex motion area will take a long time to blend into the background, resulting in a false foreground target.

[0068] S6b uses a mechanism to improve the convergence rate of the mixed Gaussian model. For each Gaussian distribution, its historical matching number is introduced into the calculation of the model learning rate.

[0069] At the beginning of background modeling, S6c averages the learning rate of each frame image becoming the background, and when the total number of video frames is less than 1 / α, each frame updates each distribution in the Gaussian mixture model with the same time weight, avoiding the first frame image from having too high a weight in the mixed Gaussian background model.

[0070] S6d In order to avoid the newly added Gaussian distribution being deleted as a low-probability event, affecting the normal background modeling and reducing the robustness of the system, a model for adaptively updating the number of Gaussian distributions is proposed. When deleting the Gaussian distribution, the average weight of the Gaussian distribution and the number of frames from the latest matching frame of a Gaussian distribution to the current frame are taken into account.

[0071] The high-altitude object dropping detection and positioning coordinates based on the improved Gaussian mixture model of the present invention are processed by grayscale image processing of the image of the building in the video frame captured by the camera, Gaussian model statistics are performed on the pixels, the next frame of the image is intercepted, and the difference of the Gaussian function of each pixel with respect to the background model is analyzed. If the match is successful, the pixel is set to 0, otherwise it is considered to be a foreground pixel. The background pixel statistical model is updated so that each frame of the image can be background-fused. Continue with the next frame of the image, and finally obtain all the coordinates of the moving object in the video. The algorithm solves the interference of diversified dynamic scenes and can effectively reduce the false alarm rate of high-altitude object dropping detection.

Claims

1. A real-time detection method for high-altitude projectiles based on an improved Gaussian mixture algorithm, characterized in that Here are the steps: Step 1: Use the Gaussian function of each pixel point of the previous N frames to establish a Gaussian mixture background statistical model: where μ k is the mean, D is a constant parameter, x is the pixel value of each image, Σ k =σ k 2 I is the covariance of the kth component, I is the RGB color space, η(x; θ k ) is the kth Gaussian normal distribution; k distributions are distributed according to the fitness value w k / σ k ,w k is the kth Gaussian component weight coefficient, σ k is the mean square error; Step 2: The first frame image in the video after the first N frames, that is, the N+1th frame image, is gray-processed according to the adaptive threshold to convert the RGB image into a gray-scale image; The algorithm is divided into three channels and runs in three parallel processes; the final partition F(xi) comes from the three parts F R (x i ),F B (x i ),F G (x i ) are operated bit by bit or one by one to obtain the required grayscale value; Step 3: Compare the background statistical model in step 1 with the Gaussian mixture model of the N+1th frame image, and add a dynamic background to the N+1th frame image: An adaptive learning rate is introduced to reflect the speed at which the foreground image merges into the background, and the estimate of t samples is recorded as Recursively write the estimation function For the t-1 sample and the ownership coefficient The last sample: Use adaptive learning rate to classify frequently moving foreground as dynamic background; Using Dynamic Estimator Make adjustments Get the adaptive threshold; Where T(x i ) is a dynamic learning controller, T inc ,T dec is a constant parameter; Step 4: Extract each pixel in the N+1th frame again, and use the Bayesian decision rule and likelihood function Analyze the difference between the Gaussian mixture model parameters of the pixel and the background model, and set the threshold; Among them: BG is the background, FG is the foreground, is the background or foreground probability at x pixels; Step 5: Set the pixel within the threshold to 0, which is the background pixel; otherwise, set the pixel to 255, which is the foreground pixel; Step 6: Update the Gaussian Mixture Model Adopting the adaptive Gaussian mixture model, the Gaussian distribution parameters of the foreground pixels, namely the mean, standard deviation, and weight, are transformed into Added to the background statistics model: And update according to the following update equation: in: is the estimated mean, is the estimated variance, used to describe the Gaussian component; is the mixing weight, and is non-negative and adds up to 1; If a pixel only appears once, reduce its weight. When the weight is less than the threshold, delete it from the mixed Gaussian model. Step 7: Continue to read the next frame of video, repeat steps 2 to 6, and finally obtain the foreground pixels of all frames of video; finally, return the selected target to the video to achieve real-time target tracking of high-altitude objects.

2. According to claim 1, the real-time detection method for high-altitude projectiles based on the improved Gaussian mixture algorithm is characterized by: The specific steps of step 6 "updating the background mixed Gaussian model" are: S6a introduces the adaptive model parameter learning rate α and sets the value of the learning rate; S6b adopts a mechanism to improve the convergence rate of the mixed Gaussian model. For each Gaussian distribution, its historical matching number is introduced into the calculation of the model learning rate; At the beginning of background modeling, S6c averages the learning rate of each frame image as the background, and when the total number of video frames is less than 1 / α, each frame updates each distribution in the Gaussian mixture model with the same time weight, avoiding the first frame image from having too high a weight in the mixed Gaussian background model; In order to avoid the newly added Gaussian distribution being deleted as a low-probability event, which affects the normal background modeling and reduces the robustness of the system, S6d adopts a model for adaptively updating the Gaussian score. When deleting the Gaussian distribution, the average weight of the Gaussian distribution and the number of frames from the latest matching frame of a Gaussian distribution to the current frame are taken into account.

3. The method for real-time detection of high-altitude projectiles based on the improved Gaussian mixture algorithm according to claim 2 is characterized in that: The value of the S6a learning rate is related to the speed at which the foreground image merges into the background. When the learning rate is large, the speed at which the foreground image merges into the background is greater than when the learning rate is small.

Images