Faster-Vibe and classification algorithm to achieve target recognition method

Abstract

The present invention belongs to the field of target detection and recognition, and specifically is a method for realizing target recognition by Faster-Vibe and classification algorithm. The method comprises: partitioning the input image according to the grayscale difference of pixels; wherein, the area with large grayscale difference has large size, and the area with small grayscale difference has small size; eliminating the area with small grayscale difference; the input image may include people, cars, small airplanes and large airplanes; by performing uneven downsampling operation on the uneliminated area, the corresponding downsampling partition is obtained; based on the Vibe algorithm, it is identified whether the pixel of each downsampling partition is a foreground pixel; the continuous foreground pixels on each downsampling partition are regarded as a pixel set of a suspected target; according to the pixel set of the suspected target, multiple holes of the target are processed, and the suspected targets are merged into the same target; the pixel set of the same target is input into a pre-trained lightweight classification network to determine which type of the target is a person, a car, a small airplane and a large airplane.

Description

Technical Field

[0001] The invention belongs to the field of target detection and recognition, and specifically relates to a target recognition method using Faster-Vibe and classification algorithms. Technical Background

[0002] In the application of airport video surveillance, it is necessary to monitor specific targets in the video in real time. Since the scale and angle of the monitored targets in the video will change greatly, especially when multiple targets appear, relying on manual monitoring of all moving objects in the video screen will not only fail to monitor every moving object, but also consume a lot of manpower, time and energy.

[0003] Analysis of existing moving target detection algorithms:

[0004] 1) The frame difference method requires a lot of calculations, and holes may appear in the image;

[0005] 2) Gaussian background modeling consumes a lot of system resources. When a sudden change in illumination occurs, it is easy to cause large-area false detection. When a static object for a long time turns into motion, it is easy to produce a "ghosting" phenomenon.

[0006] 3) The Vibe algorithm may have ghosting problems, shadow problems, and incomplete moving target problems;

[0007] 4) The biggest drawback of the moving target is that it can detect all moving objects, but the specific attributes of the objects (people, cars, airplanes, etc.) are unknown. Summary of the invention

[0008] Purpose of the invention:

[0009] In view of the problems existing in the existing target detection methods, a Faster-Vibe and classification algorithm is proposed to realize the target recognition method. By combining the traditional detection method (Vibe) with the deep learning method (ShuffleNet v2), the traditional method has small amount of calculation and occupies little memory, and the deep learning method can adaptively learn the target features, combining the advantages of both, the purpose of fast detection and recognition of targets is achieved.

[0010] Technical solution:

[0011] The present invention provides a method for realizing target recognition by using Faster-Vibe and classification algorithm, comprising:

[0012] The input image is partitioned according to the grayscale difference of pixels; the area with large grayscale difference is large in size, and the area with small grayscale difference is small in size; suppose the original image is evenly divided into n areas rect1, rect2, ...rectn, and each area is further divided into 4 areas in the form of a network. The grayscale value of each small area is counted according to the grayscale histogram, and the grayscale variation range is 0 to 255. The grayscale standard deviation of each small area is calculated, and the area with the smallest standard deviation is discarded; the input image may include people, cars, small aircraft and large aircraft;

[0013] By performing uneven downsampling operations on the non-eliminated areas, corresponding downsampling partitions are obtained;

[0014] Based on the Vibe algorithm, identify whether the pixel of each downsampled partition is a foreground pixel;

[0015] The continuous foreground pixels on each down-sampled partition are regarded as a pixel set of a suspected target;

[0016] According to the pixel set of the suspected target, multiple holes of the target are processed and the suspected targets are merged into the same target;

[0017] A set of pixels of the same object is input into a pre-trained lightweight classification network to determine whether the object is a person, a car, a small airplane, or a large airplane.

[0018] Furthermore, according to the pixel set of the suspected target, multiple holes of the target are processed and the suspected targets are merged into the same target, including:

[0019] Determine the spacing between pixel sets of all downsampled partitions;

[0020] If there are n intervals that are smaller than the preset distance, it is confirmed that the pixel sets corresponding to the n intervals are pixel sets of the same target.

[0021] Furthermore, based on the Vibe algorithm, it is identified whether the pixel of each downsampled partition is a foreground pixel, including:

[0022] For a downsampled partition, it is assumed that the pixel values ​​of each pixel and its neighboring pixels have similar distribution in the spatial domain; when the first pixel is input, that is, when t = 0, the background model of the pixel is:

[0023] BK M o =f 0 (x i ,y i )|(x i ,y i )∈N G (,y);

[0024] Among them, NG (x, y) represents the adjacent pixel values ​​in the spatial domain, f 0 (x i ,y i ) represents the pixel value of the current point. During the initialization process, N G The pixel point (x) in (x, y) i ,y l ) can be selected for L=1, 2, 3, ..., N.

[0025] (3) Perform foreground object segmentation on subsequent image sequences. When t = k, the background model of the pixel point (x, y) is The pixel value is f k (x, y). Determine whether the pixel value is foreground as follows.

[0026]

[0027] Among them, τ is randomly selected and T is the set threshold;

[0028] (4) Background model update;

[0029] Furthermore, the background model updating method includes:

[0030] Memoryless update strategy: Each time it is determined that the background model of a pixel needs to be updated, a sample value of the pixel sample set is randomly replaced by a new pixel value.

[0031] Furthermore, the background model updating method includes:

[0032] Time sampling update strategy: Update the background model at a certain update frequency. When a pixel is determined to be background, it has a probability of 1 / rate to update the background model. Rate is the time sampling factor;

[0033] Furthermore, the background model updating method includes:

[0034] Spatial neighborhood update strategy: For the pixel that needs to be updated, randomly select a background model in the neighborhood of the pixel, and update the selected background model with the new pixel.

[0035] Furthermore, the lightweight classification network is an ImageNet model, and the network is trained through image samples and target samples.

[0036] Furthermore, suppose the original image is evenly divided into n regions rect1, rect2, ... rectn, and each region is further divided into 4 regions in a network form. The grayscale value of each small region is counted according to the grayscale histogram. The grayscale variation range is 0 to 255. The grayscale standard deviation of each small region is calculated, and the region with the smallest standard deviation is discarded to perform partitioning.

[0037] The advantages of the present invention are:

[0038] Perform uneven downsampling on the image to reduce the amount of analysis and calculation and improve the accuracy of detection;

[0039] The initialization of moving targets, segmentation of backgrounds, and updating of backgrounds are all based on downsampled images, so the calculation speed is improved;

[0040] Merge the modeled targets to make the target frame more accurate and reduce the interference of miscellaneous frames;

[0041] By classifying moving targets through a classification network, targets can be distinguished from each other and the management of targets appearing in the video can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart of the method for realizing target recognition by using Faster-Vibe and classification algorithm proposed by the present invention;

[0043] Figure 2 This is the ShuffleNet v2 network training flowchart. DETAILED DESCRIPTION

[0044] The present invention is described in detail with reference to the accompanying drawings:

[0045] The present invention provides a method for realizing target recognition by using a Faster-Vibe and a classification algorithm, such as Figure 1 As shown, including:

[0046] (1) By performing uneven downsampling on the input image, the number of pixels in the image is reduced while the target features are not lost, which reduces the amount of calculation and improves the detection speed. Assume that the original image is divided into n regions rect1, rect2, ..·, rect n , after downsampling, the corresponding area is down_rect1, down_rect2, ..., down_rect n , the corresponding area scale coefficients are s1, s2, ..., s n ;

[0047] (2) Initialize the non-uniform downsampled image: Assume that the pixel values ​​of each pixel and its neighboring pixels have similar distribution in the spatial domain. When the first frame image is input, that is, at t = 0, the background model of the pixel is:

[0048] BK M o =f 0 (x i ,y i)|(x i ,y i )∈N G (x, y)

[0049] Among them, N G (x, y) represents the adjacent pixel values ​​in the spatial domain, f 0 (x i ,y i ) represents the pixel value of the current point. During the initialization process, N G The pixel point (x) in (x, y) i ,y i ) can be selected for L=1, 2, 3, ..., N.

[0050] (3) Perform foreground object segmentation on subsequent image sequences. When t = k, the background model of the pixel point (x, y) is The pixel value is f k (x, y). Determine whether the pixel value is foreground as follows.

[0051]

[0052] Among them, τ is randomly selected and T is the set threshold.

[0053] (4) Background model update. The background model can be updated using the following three methods:

[0054] 1) Memoryless update strategy: Every time it is determined that the background model of a pixel needs to be updated, a sample value of the pixel sample set is randomly replaced by a new pixel value.

[0055] 2) Time sampling update strategy. Update the background model at a certain update frequency. When a pixel is determined to be background, it has a probability of 1 / rate to update the background model. Rate is the time sampling factor, which is generally 16.

[0056] 3) Spatial neighborhood update strategy: For the pixel point that needs to be updated, a background model in the neighborhood of the pixel point is randomly selected, and the selected background model is updated with the new pixel point.

[0057] (5) Further processing is performed on multiple holes in the target after processing: the Euclidean distance between the center points of each target and the area size of the target are calculated, and the threshold c_threshold of the prior Euclidean distance is set. When the Euclidean distance is less than c_threshold, they are merged into one target.

[0058] (6) Classify the detected targets through the classification network: Considering real-time performance and accuracy, the lightweight classification network ShuffleNet v2 is used to classify the targets. The ShuffleNet v2 network training process is as follows:

[0059] 1) Initialize the network weights. Generally, the ImageNet model is used to get a good initial value.

[0060] 2) The input data is forward propagated through the convolutional layer, downsampling layer, and fully connected layer to obtain the output value;

[0061] 3) Find the error between the network output value and the target value;

[0062] 4) When the error is greater than our expected value, the error is passed back to the network, and the errors of the fully connected layer, downsampling layer, and convolutional layer are obtained in turn. The error of each layer can be understood as how much the network should bear for the total error of the network; when the error is equal to or less than our expected value, the training ends;

[0063] 5) Update the weights according to the obtained error. Then proceed to step 2).

[0064] (7) Finally, the target obtained in the downsampled image is multiplied by the corresponding ratio s n The coordinates of the target in the original image are obtained, thereby outputting the location information and attribute information of the target.

[0065] like Figure 1 As shown, the present invention is a method for realizing target recognition by using Faster-Vibe and classification algorithm, and the specific steps are as follows:

[0066] (1) By performing uneven downsampling on the input image, the number of pixels in the image is reduced while the target features are not lost, which reduces the amount of calculation and improves the detection speed. Assume that the original image is divided into n regions of different sizes rect1, rect2, ..., rect n , after downsampling, the corresponding area is down_rect1, down_rect2, ..., down_rect n , the corresponding area scale coefficients are s1, s2, ..., s n ; The area is divided according to the grayscale difference of the pixels. The larger the grayscale difference, the larger the area size, and the smaller the grayscale difference, the smaller the area size.

[0067] (2) Initialize the unevenly downsampled image: Assume that the pixel values ​​of each pixel and its neighboring pixels have similar distribution in the spatial domain. When the first frame image is input, that is, at t = 0, the background model of the pixel is:

[0068] BKM o =f 0 (x i ,y i )|(x i ,y i )∈N G (x, y);

[0069] Among them, N G (x, y) represents the adjacent pixel values ​​in the spatial domain, f 0 (x i ,y i ) represents the pixel value of the current point. During the initialization process, N G The pixel point (x) in (x, y) i ,y i ) can be selected for L=1, 2, 3, ..., N.

[0070] (3) Perform foreground object segmentation on subsequent image sequences. When t = k, the background model of the pixel point (x, y) is The pixel value is f k (x, y). Determine whether the pixel value is foreground as follows.

[0071]

[0072] Among them, τ is randomly selected and T is the set threshold.

[0073] (4) Background model update. The background model can be updated using the following three methods:

[0074] 1) Memoryless update strategy: Every time it is determined that the background model of a pixel needs to be updated, a sample value of the pixel sample set is randomly replaced by a new pixel value.

[0075] 2) Time sampling update strategy. Update the background model at a certain update frequency. When a pixel is determined to be background, it has a probability of 1 / rate to update the background model. Rate is the time sampling factor, which is generally 16.

[0076] 3) Spatial neighborhood update strategy: For the pixel point that needs to be updated, a background model in the neighborhood of the pixel point is randomly selected, and the selected background model is updated with the new pixel point.

[0077] (5) Further processing is performed on multiple holes in the target after processing: the Euclidean distance between the center points of each target and the area size of the target are calculated, and the threshold c_threshold of the prior Euclidean distance is set. When the Euclidean distance is less than c_threshold, they are merged into one target.

[0078] (6) Classify the detected target through the classification network: Considering the real-time performance and accuracy, the lightweight classification network ShuffleNet v2 is used to classify the target. The ShuffleNet v2 network training process is as follows: Figure 2 As shown, the specific steps are as follows:

[0079] 1) Initialize the network weights. Generally, the ImageNet model is used to get a good initial value.

[0080] 2) The input data is forward propagated through the convolutional layer, downsampling layer, and fully connected layer to obtain the output value;

[0081] 3) Find the error between the network output value and the target value;

[0082] 4) When the error is greater than our expected value, the error is passed back to the network, and the errors of the fully connected layer, downsampling layer, and convolutional layer are obtained in turn. The error of each layer can be understood as how much the network should bear for the total error of the network; when the error is equal to or less than our expected value, the training ends;

[0083] 5) Update the weights according to the obtained error. Then proceed to step 2).

[0084] (7) Finally, the target obtained in the downsampled image is multiplied by the corresponding ratio s n The coordinates of the target in the original image are obtained, thereby outputting the location information and attribute information of the target.

[0085] In summary, the present invention uses the Faster-Vibe algorithm to perform uneven downsampling operations on images, thereby reducing the amount of analysis and calculation and improving the accuracy of detection. At the same time, the initialization, background segmentation, and background update of moving targets are all based on downsampled images, so the calculation speed is improved. The method merges the targets after modeling to make the target frame more accurate and reduce the interference of miscellaneous frames. The mobile targets are classified through a classification network, so that the targets can be distinguished, and the targets appearing in the video can be managed in a targeted manner.

Claims

1. A method for realizing target recognition using Faster-Vibe and classification algorithm, characterized in that: include: The input image is partitioned according to the grayscale difference of pixels; the area with large grayscale difference is large in size, and the area with small grayscale difference is small in size; the original image is evenly divided into n areas rect1, rect2, ... rectn, and each area is further divided into 4 areas in the form of a network. The grayscale value of each small area is counted according to the grayscale histogram, and the grayscale variation range is 0 to 255. The grayscale standard deviation of each small area is calculated, and the area with the smallest standard deviation is discarded; the input image includes people, cars, small aircraft and large aircraft; the corresponding downsampling partitions are obtained by performing uneven downsampling operations on the areas that are not eliminated; Based on the Vibe algorithm, identify whether the pixel of each downsampled partition is a foreground pixel; The continuous foreground pixels on each down-sampled partition are regarded as a pixel set of a suspected target; According to the pixel set of the suspected target, multiple holes of the target are processed and the suspected targets are merged into the same target; Input a set of pixels of the same object into a pre-trained lightweight classification network to determine whether the object is a person, a car, a small aircraft, or a large aircraft; According to the pixel set of the suspected target, multiple holes of the target are processed and the suspected targets are merged into the same target, including: Determine the spacing between pixel sets of all downsampled partitions; If there are m intervals that are smaller than the preset distance, it is confirmed that the pixel sets corresponding to the m intervals are pixel sets of the same target.

2. The method according to claim 1, characterized in that The lightweight classification network is the ImageNet model, and the network is trained using image samples and target samples.

Images