Multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC

Abstract

The present invention discloses a multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC. The crowd segments of dense crowd images are often a large connected area. The crowd image can be preprocessed by a superpixel (SLIC) algorithm to segment the image into "granules" with similar properties, and the whole block of granules can be inferred based on granular level features. The ideal situation of this method is that the boundary between the crowd and the background area is also the boundary between different granules. For the same image, the processing effect of fine-grained images with a large number of granules and a small scale at the boundary between the crowd and the background area is more beneficial to the final segmentation result; while "granules" with a small number of granules and a large scale are more conducive to crowd feature extraction and classification. We fuse granular structure images with different granularity levels to take advantage of both. Finally, feature extraction and classification are performed on each granule, and the classification results of all granules are integrated into the original image, thereby achieving the segmentation of the crowd and the background.

Description

Technical Field

[0001] The present invention relates to the field of image analysis and understanding, and in particular to a multi-granularity fusion dense crowd segmentation algorithm based on granular computing and superpixel (SLIC) image preprocessing method. Technical Background

[0002] Large public events such as concerts, parades and rallies are common in cities around the world, and surveillance systems are an effective way to maintain safety and order in public places. With the rapid development of urbanization around the world, the formation of large crowds is becoming a norm, for example, crowds at train stations during rush hour. Therefore, crowd analysis has become a focus in the field of video surveillance to provide disaster prediction and timely support for crowd management.

[0003] In crowd analysis, crowd segmentation is usually one of the basic steps for more advanced analysis such as crowd density estimation, crowd behavior analysis, and pedestrian tracking. Before visual tasks such as tracking or behavior understanding, crowd segments need to be localized first.

[0004] Dense crowd segmentation is a challenging technical task for the following reasons:

[0005] (1) The crowd in the scene varies greatly due to crowd crowding, lighting conditions, mutual occlusion, clothing and posture differences;

[0006] (2) The background area in the dense crowd scene is cluttered with buildings, grass, and trees, which is very similar to the crowd area;

[0007] (3) The environmental layout on which the crowd formation in the scene depends varies greatly.

[0008] Due to the above problems, crowd segmentation methods for detecting each person in a crowd are still in their infancy. Some existing methods infer crowd regions by dividing the image into regular grids or overlapping multi-scale pixel grids and then extracting the pixel texture of the regular grid. Although the results are good, using pixel grids imposes certain constraints on inferring crowd segments. Crowd images are divided into regular pixel grids, where it is difficult to achieve boundary optimization of crowd segments in different scenes.

[0009] The human mind can effortlessly segment visual information with vivid details into meaningful analysis units. Transferring this structured human cognitive problem-solving ability to a crowd segmentation system can reduce the complexity of inferring crowd regions. Granular computing (GrC) is an emerging processing method in the field of information processing that simulates the human cognitive process by abstracting basic details at different granularities. By utilizing the correlation between image particles at different granularity levels, a more reasonable granularity is constructed at the boundary area between the crowd and the background, which can detect the crowd area more accurately. Summary of the invention

[0010] The purpose of the present invention is to provide a multi-granularity fusion dense crowd image segmentation algorithm based on granular computing and superpixel (SLIC) image preprocessing method to solve the problem of crowd area positioning in complex large-scale dense crowd scenes. The specific technical solution is as follows:

[0011] A multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC includes the following steps:

[0012] Step 1. Generate particle structure images at different particle size levels using the SLIC algorithm;

[0013] Step 2. Extract the particle level LRI distribution entropy of the particle structure image at different particle size levels;

[0014] Step 3. Fusion of particle structure images at different particle size levels;

[0015] Step 4. Extracting granular-level crowd image features from the fused granular structure image;

[0016] Step 5. Use the SVM classification algorithm to classify the image particles into two categories;

[0017] Step 6. Segment the crowd image.

[0018] Further, the step 1 is specifically as follows:

[0019] Step 1.1, using the SLIC algorithm to generate a coarse-grained particle structure image and a fine-grained particle structure image, the coarse-grained image is the particle structure image generated when K=100, and the fine-grained image is the particle structure image generated when K=200. The particle size level is calculated by the following formula:

[0020]

[0021] Among them, G is the granularity level, N is the total number of pixels in the image, K is the total number of particles generated in an image, and K is the parameter of the SLIC algorithm.

[0022] Further, the step 2 is specifically as follows:

[0023] Step 2.1. Select a 5×5 window as the neighborhood of the pixel point to calculate the LRI feature value of the image pixel level and obtain the LRI feature map of the image;

[0024] Step 2.2. The value range of LRI is 0-255, which is divided into 8 intervals. The LRI values ​​of all pixels in each particle of the image are distributed in the 8 intervals to obtain the frequency of pixels in each interval. The frequency of pixels in each interval is divided by the total number of pixels in the particle to obtain the frequency of occurrence of pixels in each interval. The LRI distribution entropy is then calculated using the Shannon information entropy formula.

[0025] The formula for LRI distribution entropy is described as follows:

[0026]

[0027] Where H(X) represents the value of the LRI distribution entropy of a particle, i = 1, 2, ..., 8; P i Represents the frequency of occurrence of pixels in each interval of the LRI value; extract the LRI distribution entropy features of the particle structure images at the two particle size levels of K=100 and K=200 in step 1.1 respectively.

[0028] Further, the details of step 3 are as follows:

[0029] The particle structure images generated by K = 100 and K = 200 are fused by using the LRI distribution entropy characteristics and adding the particle size and distance restrictions; K is a parameter of the SLIC algorithm, which represents the target number of particles generated by an image; the particle structure images of K = 100 and K = 200 are recorded as img1 and img2 respectively;

[0030] The specific steps of the fusion algorithm are as follows:

[0031] (1) Load the LRI distribution entropy features of img1 and img2 extracted in step 1, and set a threshold as the mean of all particle entropy values ​​in img1, denoted as Th. If the entropy values ​​of particles in img1 and img2 are both greater than or less than the threshold, they are marked as “same features”, otherwise they are marked as “different features”;

[0032] (2) Load the label matrices of img1 and img2, and find the position coordinate sets corresponding to each particle label of the two images, save them separately, and record the largest label in img1 as ml;

[0033] (3) Use the position coordinate set of a single particle in img2 obtained in the previous step to find the intersection with the position coordinate set of each particle in img1, and calculate the ratio of the intersection to the corresponding particle in img1 r1 and the ratio of the intersection to the current particle in img2 r2, and record the label of the particle in img1 that intersects with the particle in img2 as dist;

[0034] (4) Determine whether r2>=0.8 and 0.4<=r1<=0.6 are both true. If so, compare the features of the current grains in img1 and img2. If the features are the same, change the labels of the regions in img1 that overlap with all the position coordinates corresponding to the current grain in img2 to dist; if the features are different, change the labels of the regions in img1 that overlap with all the position coordinates of the current grain in img2 to ml+1, and ml=ml+1; if not, proceed to the next step.

[0035] (5) Determine whether r2>=0.8. If so, change the labels of all regions in img1 that overlap with the position coordinates corresponding to the current particle in img2 to dist; if not, proceed to the next step;

[0036] (6) Determine whether 0.4<=r1<=0.6 holds true. If so, compare the features of the current grains in img1 and img2. If the features are the same, change the labels of the regions in img1 that overlap with the current grains in img2 to dist; if the features are different, change the labels of the regions in img1 that overlap with the current grains in img2 to ml+1, with ml=ml+1. If not, do not make any changes.

[0037] (7) Repeat (3)-(6) until all particles in img2 have been judged.

[0038] Further, the details of step 4 are as follows:

[0039] Step 4.1 first extracts pixel-level LRI and LBP features;

[0040] Step 4.2: Take the LRI mean of all pixels in each particle as a feature of the particle, denoted as f1;

[0041] Step 4.3 uses the same method to extract the LRI distribution entropy feature of the fused particle structure image, denoted as f2.

[0042] Step 4.4 is similar to step 4.2. The mean of the LBP values ​​of all pixels in each particle is taken as another feature of the particle, denoted as f3.

[0043] Step 4.5: The intensity range of each color component is 0-255. Each color component is divided into 8 intervals. The frequency of pixels of each particle in the 8 intervals is counted and normalized. Finally, an 8-dimensional vector is obtained on each of the three components. The three vectors are merged together to finally obtain a 24-dimensional vector as the color histogram feature of the particle, denoted as f4.

[0044] Further, the details of step 5 are as follows:

[0045] Preprocess the features extracted in step 4, standardize the individual features, and then merge f1, f2, f3, and f4 to form a feature vector of a particle; denoted as:

[0046] f=[f1,f2,f3,f4]

[0047] Take f and the corresponding true value label as the input of the classifier and output the classification result of the corresponding particle.

[0048] Furthermore, the details of step 6 are as follows:

[0049] Find the coordinate positions of all particles in the original image, mark the classification result label of each particle to all the pixels of the particle, and obtain the segmentation result image.

[0050] The multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC of the present invention, on the basis of constructing granular images of different granularity levels for images by the SLIC algorithm, utilizes the correlation between the LRI distribution entropy of the particles and the color features to fuse the granular images of different granularity levels, thereby absorbing the advantages of images of different granularity levels and better locating crowd segments. The method proposed by the present invention is novel, simple and easy to understand, and has been verified to be effective by examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the algorithm flow chart of the present invention;

[0052] Figure 2a , Figure 2b It is the particle structure image at different particle size levels of the present invention;

[0053] Figure 3a , Figure 3b It is the LRI distribution entropy characteristic diagram of the particle structure image of the present invention;

[0054] Figure 4 This is the particle structure image after the different particle size levels are fused according to the present invention.

[0055] Figure 5 This is the final pixel-level crowd segmentation result map. DETAILED DESCRIPTION

[0056] The present invention will be further described below in conjunction with the accompanying drawings.

[0057] The technical solution adopted by the present invention is: first, the SLIC superpixel algorithm is used to divide the image into individual "grains" at different granularity levels. Ideally, the boundaries of these grains are just the boundaries between the crowd and the background area. In order to get closer to this goal, the "grain" images of different granularity levels of the same image are used for fusion to generate a more appropriate grain image. This "appropriate" grain image can be closer to the real boundary of the crowd scene at the boundary, and can more effectively extract grain-level features.

[0058] like Figure 1 As shown, the multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC of the present invention includes the following steps:

[0059] Step 1. Generate particle structure images at different particle size levels using the SLIC algorithm;

[0060] Step 2. Extract the particle level LRI distribution entropy of the particle structure image at different particle size levels;

[0061] Step 3. Fusion of particle structure images at different particle size levels;

[0062] Step 4. Extracting granular-level crowd image features from the fused granular structure image;

[0063] Step 5. Use the SVM classification algorithm to classify the image particles into two categories;

[0064] Step 6. Segment the crowd image.

[0065] The step 1 is specifically as follows:

[0066] Step 1.1, using the SLIC algorithm to generate a coarse-grained particle structure image and a fine-grained particle structure image, the coarse-grained image is the particle structure image generated when K=100, and the fine-grained image is the particle structure image generated when K=200. The particle size level is calculated by the following formula:

[0067]

[0068] Among them, G is the granularity level, N is the total number of pixels in the image, K is the total number of particles generated in an image, and K is the parameter of the SLIC algorithm.

[0069] The step 2 is specifically as follows:

[0070] Step 2.1, LRI (Local Range of Intensity) is a local intensity range feature of grayscale images, which is defined as the difference between the maximum intensity and the minimum intensity of the local neighborhood of the pixel of interest. We select a 5×5 window as the neighborhood of the pixel point to calculate the LRI feature value at the pixel level of the image and obtain the LRI feature map of the image.

[0071] Step 2.2, the value range of LRI is 0-255, which is divided into 8 intervals. The LRI values ​​of all pixels in each particle of the image are distributed in the 8 intervals to obtain the pixel frequency in each interval. The pixel frequency in each interval is divided by the total number of pixels in the particle to obtain the frequency of occurrence of the pixel in each interval. The LRI distribution entropy is then calculated using the Shannon information entropy formula.

[0072] The formula for LRI distribution entropy is described as follows:

[0073]

[0074] Where H(X) represents the value of the LRI distribution entropy of a particle. i=1,2,...,8; P i Indicates the frequency of occurrence of pixels in each interval of LRI value. We extract LRI distribution entropy features for the particle structure images at two particle size levels of K = 100 and K = 200 in step 1.1.

[0075] Due to the changes in the appearance of different individuals in a dense crowd, the crowd part often shows a larger range of intensity changes than the background (non-crowd) area. Most of the background area has a small LRI value, and the LRI value distribution is relatively simple, so the LRI distribution entropy of the background area particles is small; the LRI value of the crowd area is large, while the LRI value of the part of the person that is not at the edge is small. The boundary area between the crowd and the background will have a larger LRI value distribution because it is at the edge, so the LRI value of the crowd area and the boundary will be more evenly distributed, and the LRI distribution entropy is larger.

[0076] Step 3 is as follows:

[0077] Step 2 extracts the LRI distribution entropy features of the particles at different particle size levels. We use this feature, combined with particle size and distance constraints, to fuse the particle structure images generated by K = 100 and K = 200. K is a parameter of the SLIC algorithm, which represents the target number of particles generated by an image. The particle structure images with K = 100 and K = 200 are recorded as img1 and img2 respectively.

[0078] The specific steps of the fusion algorithm are as follows:

[0079] (1) Load the LRI distribution entropy features of img1 and img2 extracted in step 1, and set a threshold as the mean of all particle entropy values ​​in img1, denoted as Th. If the entropy values ​​of particles in img1 and img2 are both greater than or less than the threshold, they are marked as “same features”, otherwise they are marked as “different features”;

[0080] (2) Load the label matrices of img1 and img2, and find the position coordinate sets corresponding to each particle label of the two images, save them separately, and record the largest label in img1 as ml;

[0081] (3) Use the position coordinate set of a single particle in img2 obtained in the previous step to find the intersection with the position coordinate set of each particle in img1, and calculate the ratio of the intersection to the corresponding particle in img1 r1 and the ratio of the intersection to the current particle in img2 r2, and record the label of the particle in img1 that intersects with the particle in img2 as dist;

[0082] (4) Determine whether r2>=0.8 and 0.4<=r1<=0.6 are both true. If so, compare the features of the current grains in img1 and img2. If the features are the same, change the labels of the regions in img1 that overlap with all the position coordinates corresponding to the current grain in img2 to dist; if the features are different, change the labels of the regions in img1 that overlap with all the position coordinates of the current grain in img2 to ml+1, and ml=ml+1; if not, proceed to the next step.

[0083] (5) Determine whether r2>=0.8. If so, change the labels of all regions in img1 that overlap with the position coordinates corresponding to the current particle in img2 to dist; if not, proceed to the next step;

[0084] (6) Determine whether 0.4<=r1<=0.6 holds true. If so, compare the features of the current grains in img1 and img2. If the features are the same, change the labels of the regions in img1 that overlap with the current grains in img2 to dist; if the features are different, change the labels of the regions in img1 that overlap with the current grains in img2 to ml+1, with ml=ml+1. If not, do not make any changes.

[0085] (7) Repeat (3)-(6) until all particles in img2 have been judged.

[0086] Step 4 is as follows:

[0087] Step 4.1: After obtaining the fused particle structure image in step 3, it is necessary to extract the particle-level crowd image features for binary classification of these particles. As needed, we need to first extract the pixel-level LRI and LBP features.

[0088] LRI (Local Range of Intensity) is a local intensity range feature of grayscale images, which is defined as the difference between the maximum intensity and the minimum intensity of the local neighborhood of the pixel of interest. We choose a 5×5 window as the neighborhood of the pixel point to calculate the LRI feature value at the pixel level of the image and obtain the LRI feature map of the image.

[0089] LBP (Local Binary Pattern) is an operator used to describe the local texture features of an image; it has significant advantages such as rotation invariance and grayscale invariance. We calculate the LBP value of each pixel and obtain the LBP feature map.

[0090] Step 4.2, generally, due to the variation in the appearance of different individuals in a dense crowd, the crowd part tends to show a larger range of intensity variation than the background (non-crowd) area, that is, the LRI value is larger. We take the LRI mean of all pixels in each grain as a feature of the grain, denoted as f1.

[0091] Step 4.3, in step 2.2, we extracted the LRI distribution entropy features of the original coarse-grained and fine-grained images, and used the same method to extract the LRI distribution entropy features of the fused particle structure image, denoted as f2.

[0092] Step 4.4, similar to step 4.2, we take the mean of the LBP values ​​of all pixels in each particle as another feature of the particle, denoted as f3.

[0093] Step 4.5, according to our observation of dense crowd images, the crowd area has a variety of colors due to the different clothes, skin color, hair color and items carried by different people, while non-crowd areas, such as buildings, grass, cars, etc., often have a single color in a local area, so the color features of the crowd and background areas at the granular level have a great degree of distinction. In the RGB color system, different colors are mixed with red component R, green component G and red component B with different intensities, so the intensity distribution of different colors in the three components is different, so the color features can be represented by the histogram of the three color components.

[0094] The intensity range of each color component is 0-255. We divide each color component into 8 intervals, count the frequency of pixels of each particle in the 8 intervals, normalize them, and finally get an 8-dimensional vector on each of the three components. The three vectors are combined to finally get a 24-dimensional vector as the color histogram feature of the particle, denoted as f4.

[0095] Step 5 is as follows:

[0096] We used 40 images to verify the method of the present invention, among which 32 images were randomly selected as training sets, all particles in the 32 images were trained, and the parameters were adjusted using the grid search method to establish a classifier model; the remaining 8 images were used as test sets, and all particles were classified using the established classifier.

[0097] The features extracted in step 4 are preprocessed, and the single features are standardized. Then f1, f2, f3, and f4 are merged to form a feature vector of a particle. Since f1, f2, and f3 are all 1-dimensional and f4 is 24-dimensional, a 27-dimensional vector is finally formed, which is recorded as:

[0098] f=[f1,f2,f3,f4]

[0099] Take f and the corresponding true value label as the input of the classifier and output the classification result of the corresponding particle.

[0100] Step 6 is as follows:

[0101] Find the coordinate positions of all particles in the original image, mark the classification result label of each particle to all the pixels of the particle, and obtain the segmentation result image.

[0102] Example

[0103] Step 1. Generate particle structure images at different particle size levels using the SLIC algorithm;

[0104] Step 1.1, such as Figure 2a The particle structure image when K = 100 is shown. Figure 2b The particle structure image when K=200 is shown.

[0105] Step 2, extracting the particle level LRI distribution entropy of the particle structure image at different particle size levels;

[0106] like Figure 3a and 3b As shown in the figure, the average entropy of all particles in a picture is used as the threshold. The area covered in red is the area where the LRI distribution entropy is less than the threshold, which is mainly the background area; the LRI distribution entropy of the boundary and the crowd area is greater than the threshold.

[0107] Step 3, fusion of particle structure images at different particle size levels;

[0108] like Figure 4 , showing the results of particle structure image fusion when K = 100 and K = 200.

[0109] Step 4, extracting granular-level crowd image features from the fused granular structure image;

[0110] Step 5, use the SVM classification algorithm to classify the image particles into two categories;

[0111] Step 6: Segment the crowd image.

[0112] like Figure 5 As shown, the pixel-level crowd segmentation results are shown.

[0113] The embodiments described above are merely descriptions of preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC, comprising the following steps: Step 1. Generate particle structure images at different particle size levels using the SLIC algorithm; Step 2. Extract the particle level LRI distribution entropy of the particle structure image at different particle size levels; Step 3. Fusion of particle structure images at different particle size levels; Step 4. Extracting granular-level crowd image features from the fused granular structure image; Step 5. Use the SVM classification algorithm to classify the image particles into two categories; Step 6. Segment the crowd image; The step 1 is specifically as follows: Step 1.1, using the SLIC algorithm to generate a coarse-grained particle structure image and a fine-grained particle structure image, the coarse-grained image is the particle structure image generated when K=100, and the fine-grained image is the particle structure image generated when K=200. The particle size level is calculated by the following formula: in, G is the granularity level, N is the total number of pixels in the image, K is the total number of particles generated in an image, and K is a parameter of the SLIC algorithm; The step 2 is specifically as follows: Step 2.

1. Select a 5×5 window as the neighborhood of the pixel point to calculate the LRI feature value of the image pixel level and obtain the LRI feature map of the image; Step 2.

2. The value range of LRI is 0-255, which is divided into 8 intervals. The LRI values ​​of all pixels in each particle of the image are distributed in the 8 intervals to obtain the frequency of pixels in each interval. The frequency of pixels in each interval is divided by the total number of pixels in the particle to obtain the frequency of occurrence of pixels in each interval. The LRI distribution entropy is then calculated using the Shannon information entropy formula. The formula for LRI distribution entropy is described as follows: Where H(X) represents the value of the LRI distribution entropy of a particle, i = 1, 2, ..., 8; P i Indicates the frequency of occurrence of pixels in each interval of LRI value; extract LRI distribution entropy features for the particle structure images at two particle size levels of K = 100 and K = 200 in step 1.1; The details of step 3 are as follows: The particle structure images generated by K = 100 and K = 200 are fused by using the LRI distribution entropy characteristics and adding the particle size and distance restrictions; K is a parameter of the SLIC algorithm, which represents the target number of particles generated by an image; the particle structure images of K = 100 and K = 200 are recorded as img1 and img2 respectively; The specific steps of the fusion algorithm are as follows: (1) Load the LRI distribution entropy features of img1 and img2 extracted in step 1, and set a threshold as the mean of all particle entropy values ​​in img1, denoted as Th. If the entropy values ​​of particles in img1 and img2 are both greater than or less than the threshold, it is recorded as "same features", otherwise it is recorded as "different features"; (2) Load the label matrices of img1 and img2, and find the position coordinate sets corresponding to each particle label of the two images, save them separately, and record the largest label in img1 as ml; (3) Use the position coordinate set of a single particle in img2 obtained in the previous step to find the intersection with the position coordinate set of each particle in img1, and calculate the ratio of the intersection to the corresponding particle in img1 r1 and the ratio of the intersection to the current particle in img2 r2, and record the label of the particle in img1 that intersects with the particle in img2 as dist; (4) Determine whether r2>=0.8 and 0.4<=r1<=0.6 are both true. If so, compare the features of the current grains in img1 and img2. If the features are the same, change the labels of the regions in img1 that overlap with all the position coordinates corresponding to the current grain in img2 to dist; if the features are different, change the labels of the regions in img1 that overlap with all the position coordinates of the current grain in img2 to ml+1, and ml=ml+1; if not, proceed to the next step; (5) Determine whether r2>=0.

8. If so, change the labels of all regions in img1 that overlap with the position coordinates corresponding to the current particle in img2 to dist; if not, proceed to the next step; (6) Determine whether 0.4<=r1<=0.6 holds true. If so, compare the features of the current grains in img1 and img2. If the features are the same, change the labels of the regions in img1 that overlap with the coordinates of all positions corresponding to the current grain in img2 to dist; if the features are different, change the labels of the regions in img1 that overlap with the coordinates of all positions corresponding to the current grain in img2 to ml+1, with ml=ml+1; if not, do not make any changes. (7) Repeat (3)-(6) until all particles in img2 have been judged; The details of step 4 are as follows: Step 4.1 first extracts pixel-level LRI and LBP features; Step 4.2: Take the LRI mean of all pixels in each particle as a feature of the particle, denoted as f1; Step 4.3 uses the same method to extract the LRI distribution entropy feature of the fused particle structure image, denoted as f2; Step 4.4 is similar to step 4.

2. The mean of the LBP values ​​of all pixels in each particle is taken as another feature of the particle, denoted as f3. Step 4.5: The intensity range of each color component is 0-255. Each color component is divided into 8 intervals. The frequency of pixels of each particle in the 8 intervals is counted and normalized. Finally, an 8-dimensional vector is obtained on each of the three components. The three vectors are merged together to finally obtain a 24-dimensional vector as the color histogram feature of the particle, denoted as f4.

2. The multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC as claimed in claim 1 is characterized by: The details of step 5 are as follows: Preprocess the features extracted in step 4, standardize the individual features, and then merge f1, f2, f3, and f4 to form a feature vector of a particle; denoted as: f=[f1,f2,f3,f4] Take f and the corresponding true value label as the input of the classifier and output the classification result of the corresponding particle.

3. The multi-granularity fusion dense crowd segmentation method based on granular computing and SLIC as claimed in claim 2 is characterized by: The details of step 6 are as follows: Find the coordinate positions of all particles in the original image, mark the classification result label of each particle to all the pixels of the particle, and obtain the segmentation result image.

Images