A camera monitoring network target matching method based on a camera network topology

Abstract

The application discloses a camera monitoring network target matching method based on a camera network topology, which comprises the following steps: S1, collecting relevant scene data, constructing a training set according to target images in the relevant scene data, training a target matching network, and obtaining a target matching model; S2, inputting a target image to be matched into the target matching model to obtain a preliminary target matching image sequence; and S3, reordering the preliminary target matching image sequence through a feature distance measurement formula based on the camera network topology to obtain a target matching result. The application reorders the preliminary target matching image sequence output by the target matching model obtained through training by using information of the camera network topology, improves the matching accuracy, and further takes into account the performance and efficiency of the target matching method under complex conditions.

Description

Technical Field

[0001] This invention belongs to the field of image target matching, specifically relating to a target matching method for camera monitoring networks based on camera network topology. Background Technology

[0002] Target matching in camera surveillance networks typically refers to monitoring various parts of a large venue, such as airport terminals, shopping malls, and campuses. In these scenarios, the surveillance networks usually consist of numerous and dispersed cameras, whose fields of view generally do not overlap. Furthermore, different cameras typically capture different scenes. For people or objects in different scenes, sometimes it's necessary to stably track or monitor specific targets throughout the entire process, requiring the matching of targets appearing in the fields of view of different cameras. This type of problem is usually reduced to automatically identifying and distinguishing a single target from multiple views obtained by the camera network.

[0003] Target matching in different scenarios is beneficial for both full-field target tracking and tracking of target movement trajectories. It also helps to uncover behavioral information in different scenarios, such as determining target behavior based on the temporal sequence of the matched scenarios—that is, the connection between the target and the scene in which it appears under the camera's view. However, target matching faces numerous challenges in large venues: a large number of cameras, numerous targets and other objects, mutual occlusion between different targets and between targets and other objects, potential changes in target posture and appearance, and changes in camera perspective and lighting. Therefore, target matching faces significant spatiotemporal ambiguity and high complexity, and the non-overlapping fields of view of multiple cameras further complicates the process.

[0004] Traditional target matching methods are mostly based on the appearance of the target, that is, by extracting the appearance features of the target in the image for matching. However, in complex situations, in order to extract effective features, complex image processing is required. Although this method improves accuracy, it is too time-consuming. In addition, in camera networks, targets often appear in multiple cameras, that is, the same target appears in multiple images, which increases the time cost of target matching. Therefore, it is possible to consider incorporating other information, such as the topology information of the camera network, to improve matching efficiency and accuracy. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a target matching method for camera monitoring networks based on camera network topology, which solves the problem that existing target matching methods cannot simultaneously achieve both efficiency and performance in complex situations.

[0006] To achieve the aforementioned objectives, the present invention employs the following technical solution: a target matching method for a camera monitoring network based on camera network topology, comprising the following steps:

[0007] S1. Collect relevant scene data, construct a training set based on the target images in the relevant scene data, train a target matching network, and obtain a target matching model;

[0008] S2. Input the target image to be matched into the target matching model to obtain a preliminary target matching image sequence;

[0009] S3. The initial target matching image sequence is reordered using a feature distance metric formula based on the camera network topology to obtain the target matching result.

[0010] Further: S1 includes the following sub-steps:

[0011] S11. Collect relevant scene data, select the target image from the relevant scene data, and use the random erasure method to randomly erase the target image to obtain the training set;

[0012] S12. Obtain the target matching network and train it using the training set to obtain the target matching model.

[0013] Further: In S11, the random erasure method specifically includes:

[0014] SA1: Preset random erasure probability parameters. Select the image to be randomly erased from the target image according to the random erasure probability parameters to obtain the selected image.

[0015] SA2. Select a rectangular area in the selected image to obtain the selected rectangular area. Fill the pixel values ​​of the selected rectangular area to complete the random erasure of the target image.

[0016] Further: In SA2, the method for obtaining the selected rectangular region is specifically as follows:

[0017] Select a point P in the selected image. When the coordinates of point P are (x, y) e ,y e If the following formula is satisfied, then the region I constructed by point P will be... e =(x e ,y e ,x e +W e ,y e +H e () as the selected rectangular area;

[0018] x e +W e ≤W

[0019] y e +H e ≤H

[0020] In the formula, W is the length of the selected image, H is the width of the selected image, and W... e H is the length of the selected rectangular region. e The width of the selected rectangular area;

[0021] In the SA2, the method for filling the pixel values ​​of the selected rectangular region is as follows:

[0022] The values ​​of each pixel in the selected rectangular area are filled by filling pixel values. The specific expression is:

[0023]

[0024] In the formula, p v Here, c represents the value of each pixel within the selected rectangular region, and n represents the number of channels in the image pixel. p This represents the total number of pixels in the selected rectangular area.

[0025] Further: S12 specifically includes:

[0026] Obtain the target matching network, train it using the training set, and incorporate camera network topology information for performance compensation to obtain the target matching model.

[0027] Furthermore: S2 specifically refers to:

[0028] The target image to be matched and several search images are input into the target matching model to obtain the feature distance between the target image to be matched and each search image. All search images are sorted in ascending order according to the feature distance between the target image to be matched and each search image to obtain a preliminary target matching image sequence.

[0029] The retrieved image is obtained by searching the database for the target image to be matched.

[0030] Further: In S2, the target image p to be matched and the i-th searched image g i Feature distance d m (p,g i The specific expression for ) is:

[0031]

[0032] In the formula, x p x is the feature vector extracted from the target image by the target matching model. giLet M be the feature vector extracted from the i-th retrieved image by the target matching model, where i = 1, 2, ..., N, N is the total number of retrieved images, and M is a positive semi-definite matrix. This is a transpose operation.

[0033] Further: S3 includes the following sub-steps:

[0034] S31. Obtain camera network distribution information, label the cameras in the network according to the camera network distribution information, and arrange each labeled camera according to the distance between the labeled cameras to obtain the camera distance sequence corresponding to each labeled camera. Obtain the camera distance sequence set based on the camera distance sequences corresponding to all labeled cameras.

[0035] S32. Based on the set of distance sequences of cameras, obtain the network topology weight of each labeled camera according to the camera distance sequence corresponding to each labeled camera, and obtain the set of network topology weights according to the network topology weights of all cameras;

[0036] S33. Add the network topology weight set to the reordering algorithm based on K nearest neighbors to obtain the feature distance metric formula based on camera network topology;

[0037] S34. The preliminary target matching image sequence is reordered according to the feature distance metric formula based on camera network topology to obtain the target matching result.

[0038] Furthermore: In S33, the feature distance metric formula based on the camera network topology... Specifically, it is as follows:

[0039]

[0040] In the formula, p cam The image of the target to be re-identified, captured by the camera cam, where cam is the camera number of the target image p to be matched, and t is the target image p to be re-identified. cam,c Let c be the network topology weight of camera c relative to camera cam, and c = 1, 2, ..., N. c N c The total number of cameras, To sum the values ​​of the camera labels, where j is the camera label, a is the topology parameter, and g is the summation parameter. i,c To retrieve the image dataset from the database, d J (p,g i,c ) for the improved Jaccard distance, d m (p,g i Let p be the target image to be matched and g be the ith search image. i The feature distance, i = 1, 2...N, where N is the total number of images retrieved.

[0041] Furthermore: the improved Jaccard distance d J (p,g i,c The specific expression for ) is:

[0042]

[0043] In the formula, The target image p to be matched is compared with the image g in the search image library. j The numerical values ​​of the K nearest neighbor images after modification are in the range of [0,1]. To retrieve image g i Image g is compared with other images in the image library and the target image p to be matched. j The value of the modified K nearest neighbor image set is quantified and its range is between [0,1].

[0044] The beneficial effects of this invention are as follows:

[0045] (1) The present invention provides a target matching method for camera monitoring network based on camera network topology to train an existing target matching model. This model does not have to be the most advanced, best performing, and large-scale model. Instead, a small model with good performance can be used, thereby reducing the cost of application and making deployment more convenient and faster.

[0046] (2) By utilizing the information of the camera network topology, the present invention reorders the preliminary target matching image sequence output by the trained target matching model, thereby improving the accuracy of matching and taking into account both the performance and efficiency of the target matching method in complex situations.

[0047] (3) The reordering of the feature distance metric formula based on the camera network topology of this invention improves the accuracy of a single matching result while providing multiple accurate matching results, which is beneficial for the mining of target movement trajectories. Attached Figure Description

[0048] Figure 1 This is a flowchart of a camera monitoring network target matching method based on camera network topology according to the present invention.

[0049] Figure 2 This is a schematic diagram of the random erasure method of the present invention.

[0050] Figure 3 This is a schematic diagram illustrating the camera network topology reordering method of the present invention. Detailed Implementation

[0051] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0052] like Figure 1 As shown, in one embodiment of the present invention, a target matching method for a camera monitoring network based on camera network topology includes the following steps:

[0053] S1. Collect relevant scene data, construct a training set based on the target images in the relevant scene data, train a target matching network, and obtain a target matching model;

[0054] S2. Input the target image to be matched into the target matching model to obtain a preliminary target matching image sequence;

[0055] S3. The initial target matching image sequence is reordered using a feature distance metric formula based on the camera network topology to obtain the target matching result.

[0056] S1 includes the following steps:

[0057] S11. Collect relevant scene data, select the target image from the relevant scene data, and use the random erasure method to randomly erase the target image to obtain the training set;

[0058] S12. Obtain the target matching network and train it using the training set to obtain the target matching model.

[0059] like Figure 2 As shown, in this embodiment, a corresponding target matching dataset or related images are collected according to the application scenario, and the target image is extracted through a target detection algorithm. Then, the target image is randomly erased using a random erasure method.

[0060] In step S11, the random erasure method is specifically as follows:

[0061] SA1, Preset random erasure probability parameter p e Based on the random erasure probability parameter, select the image that needs to be randomly erased from the target image to obtain the selected image;

[0062] SA2. Select a rectangular area in the selected image to obtain the selected rectangular area. Fill the pixel values ​​of the selected rectangular area to complete the random erasure of the target image.

[0063] In this embodiment, a preset random erasure probability parameter p is used.e That is, for an image, the probability of it being randomly erased is p. e The probability of keeping it as it is is 1-p. e

[0064] In SA2, the method for obtaining the selected rectangular region is as follows:

[0065] Select a point P in the selected image. When the coordinates of point P are (x, y) e ,y e If the following formula is satisfied, then the region I constructed by point P will be... e =(x e ,y e ,x e +W e ,y e +H e () as the selected rectangular area;

[0066] x e +W e ≤W

[0067] y e +H e ≤H

[0068] In the formula, W is the length of the selected image, H is the width of the selected image, and W... e H is the length of the selected rectangular region. e The width of the selected rectangular area;

[0069] In this embodiment, the area I to be erased is randomly erased. e Aspect Ratio H e / W e Initialize to values ​​in the range (r1, r2), and W e H e r1 and r2 are parameter values ​​that are preset based on the target image size W and H, and can be adjusted at any time during training.

[0070] In the SA2, the method for filling the pixel values ​​of the selected rectangular region is as follows:

[0071] The values ​​of each pixel in the selected rectangular area are filled by filling pixel values. The specific expression is:

[0072]

[0073] In the formula, p v Here, c represents the value of each pixel within the selected rectangular region, and n represents the number of channels in the image pixel. pThis represents the total number of pixels in the selected rectangular area. This means summing the pixel values ​​of all channels for all pixels and then dividing by the product of the number of channels and the total number of pixels to obtain the average pixel value of the image.

[0074] Specifically, S12 is:

[0075] Obtain the target matching network, train it using the training set, and incorporate camera network topology information for performance compensation to obtain the target matching model.

[0076] In this embodiment, the target matching network obtained is an existing network. A target matching network with relatively good performance and small network size can be adopted. Performance compensation is performed by adding camera network topology information. That is, introducing camera network topology information can reduce the performance requirements of the target matching network and achieve rapid deployment of target matching capabilities and cost reduction.

[0077] Specifically, S2 is:

[0078] The target image to be matched and several search images are input into the target matching model to obtain the feature distance between the target image to be matched and each search image. All search images are sorted in ascending order according to the feature distance between the target image to be matched and each search image to obtain a preliminary target matching image sequence.

[0079] The retrieved image is obtained by searching the database for the target image to be matched.

[0080] In step S2, the target image p to be matched and the i-th search image g i Feature distance d m (p,g i The specific expression for ) is:

[0081]

[0082] In the formula, x p x is the feature vector extracted from the target image by the target matching model. gi Let M be the feature vector extracted from the i-th retrieved image by the target matching model, where i = 1, 2, ..., N, N is the total number of retrieved images, and M is a positive semi-definite matrix. This is a transpose operation.

[0083] In this embodiment, the reference image dataset containing N search images retrieved from the database is G = {g i |=1,2…N}, the target image p to be matched and the i-th search image g i Feature distance d m (p,g iThe size of ) is used to determine whether the target image to be matched matches the retrieved image.

[0084] S3 includes the following steps:

[0085] S31. Obtain camera network distribution information, label the cameras in the network according to the camera network distribution information, and arrange each labeled camera according to the distance between the labeled cameras to obtain the camera distance sequence corresponding to each labeled camera. Obtain the camera distance sequence set based on the camera distance sequences corresponding to all labeled cameras.

[0086] In this embodiment, p is used cam This refers to the image of the target that needs to be re-identified, captured by the camera (cam). i,c This indicates that the image dataset G′={g} in the database is retrieved. i,c |i=1,2…N,c=1,2…N c The image in the search box is the i-th image belonging to camera c, and N is the total number of search images. c Let j be the total number of cameras. When the camera capturing the target image to be matched is labeled j, the camera labels are arranged according to the distances between other cameras and camera j, resulting in the camera distance sequence corresponding to camera j. In this algorithm, cameras with closer distances are ranked higher. Since it's possible that a camera might have multiple images of the target, c1 = the current camera's label is ranked first. For each labeled camera in the camera network, a distance sequence with other cameras can be obtained, thus creating a set of camera distance sequences.

[0087] S32. Based on the set of distance sequences of cameras, obtain the network topology weight of each labeled camera according to the camera distance sequence corresponding to each labeled camera, and obtain the set of network topology weights according to the network topology weights of all cameras;

[0088] In this embodiment, based on the camera distance sequence corresponding to camera k We obtain a camera network topology weight t for each camera relative to camera k. k =[1,2,…,N c That is, the weight of camera k to camera k is 1, and the weight of camera c2 to camera k is 2. The weight of camera k is N. c This allows us to obtain the network topology weights of each labeled camera in the camera network, thus yielding a set of network topology weights.

[0089] S33. Add the network topology weight set to the reordering algorithm based on K nearest neighbors to obtain the feature distance metric formula based on camera network topology;

[0090] S34. Reorder the preliminary target matching image sequence according to the feature distance metric formula based on camera network topology to obtain the target matching result.

[0091] In S33, the feature distance metric formula based on the camera network topology Specifically, it is as follows:

[0092]

[0093] In the formula, cam is the camera label of the target image p to be matched, and t cam,c Let c be the network topology weight of camera c relative to camera cam, and c = 1, 2, ..., N. c N c The total number of cameras, To sum the values ​​of the camera labels, where j is the camera label, a is the topology parameter, and g is the summation parameter. i,c To retrieve the image dataset from the database, d J (p,g i,c ) for the improved Jaccard distance, d m (p,g i Let p be the target image to be matched and g be the ith search image. i The feature distance, i = 1, 2...N, where N is the total number of images retrieved.

[0094] In this embodiment, the method for obtaining the improved Jaccard distance is as follows:

[0095] The Jaccard distance is improved by modifying it using the K-nearest neighbor algorithm.

[0096] The K-Nearest Neighbor (KNN) algorithm calculates the feature distance between the test image and the reference image set, and arranges the K images into a matching candidate queue in descending order. The K-Mutual Nearest Neighbor (KML) algorithm states that for two similar images A and B, if B is in A's KML queue, A should also be in B's KML queue. If A and B are not similar, then even if B is in A's KML queue, A is unlikely to appear in B's KML queue. This bidirectional verification improves the accuracy of re-ranking.

[0097] Jaccard distance is a metric for calculating the difference between two sets. It is defined as the number of elements in the intersection of the two sets divided by the number of elements in the union. In the K nearest neighbor algorithm, the set N(p,k) of the K nearest neighbor images of the target image p is first defined as follows:

[0098]

[0099] In the formula, This represents the retrieved image whose first k feature distances are calculated between the target image p and the reference image dataset G.

[0100] Secondly, the set of K mutually nearest neighbor reference images R(p, k) is defined as follows:

[0101] R(p, k) = {g i |(g i ∈N(p,k))∩(p∈N(g i ,k))}

[0102] In the formula, N(g) i (k) represents the i-th retrieved image g. i The set of the K nearest neighbor images, at which point g is retrieved. i When comparing similar images, the reference image dataset contains image p, so the set of K nearest neighbor images is N(g i The intersection of N(p,k) and N(p,k).

[0103] Due to variations in lighting conditions, pose, camera viewpoint, and occlusion, the truly matching image is likely to be removed from the K nearest neighbor reordering queue and is not included in the set of K nearest neighbors. The K nearest neighbor reordering algorithm modifies the set; in the set of K nearest neighbors, it conditionally combines half of the K nearest neighbors of each image in R(p,k) with R(p,k), as shown in the following expression:

[0104]

[0105]

[0106] In the formula, q is an image in the reference image set R(p,k). If half of the K nearest neighbors of q are... The number of intersection images with R(p,k) is greater than or equal to The number of images in the file, then... The union of R(p,k) with R(p,k) is then obtained to form a new set R. * (p, k).

[0107] Therefore, in the K-nearest neighbor algorithm, the Jaccard distance d J (p, g) i The calculation formula can be expressed as:

[0108]

[0109] Since performing intersection and union operations on image sets is too time-consuming, and the distance calculation method treats all nearest neighbor images as equally important (in reality, in most cases, images with smaller feature distances are more likely to be positive samples close to p), a K-nearest neighbor-based reordering algorithm encodes the K nearest neighbor image sets into an equivalent, structurally simple N-dimensional binary vector to save computation time. The element definition formula in a binary vector for:

[0110]

[0111] In the formula, g i ∈R * (p, k) represents g in the image reference dataset. i The modified set R of the K nearest neighbors of p * (p, k), the element at this time The value is 1 if the vector contains elements with a value of 1, and zero otherwise. Thus, the set of elements in the vector with a value of 1 represents the set of the K nearest neighbor images of p.

[0112] To account for the influence of feature distance, an exponential form was used to modify the weights. The element definition in the modified binary vector is as follows: As shown in the following formula:

[0113]

[0114] Where d(p, g) i ) are p and g i The smaller the Mahalanobis distance, the better. The closer it is to 1.

[0115] Similarly, according to V p The definition of g allows us to obtain the retrieved image g. i The vector representing the K nearest neighbor images of all images retrieved from the image library and the image to be matched (p) is modified from the vector of the image to be matched.

[0116] At this point, the number of candidate images in the intersection and complement of the image sets can be rewritten as:

[0117]

[0118]

[0119] In the formula, min and max are based on V p , The minimum and maximum values ​​are selected based on the size of the corresponding elements, and then a new vector is obtained. ||1 represents the calculation of the L1 norm of the vector, which is the sum of the absolute values ​​of each element of the vector.

[0120] Vector V p and Since all elements in the vector are greater than 0, its L1 norm is equal to the sum of all vector elements. and The summation results after comparing sizes, where p and g are the images to be matched. j This refers to images in the search database. For a specific image g in the search database... i g j The Jaccard distance, representing the search for other images or matching images p in the database, is ultimately expressed as:

[0121]

[0122] In the formula, The target image p to be matched is compared with the image g in the search image library. j The numerical values ​​of the K nearest neighbor images after modification are in the range of [0, 1]. To retrieve image g i Image g is compared with other images in the image library and the target image p to be matched. j The value of the modified K nearest neighbor image set is quantified and its range is between [0, 1].

[0123] like Figure 3 As shown in the figure, in this embodiment, the camera network has a total of 8 cameras. Assuming that the camera number of the image to be matched is 4, its corresponding camera distance sequence s4 = [4, 3, 5, 6, 7, 2, 8, 1], and the corresponding camera weight matrix is ​​t. k = [1, 2, ..., 8]; Assuming the camera number of the retrieved image is 2, its corresponding weight is 6. The feature distance between the two images is then:

[0124]

[0125] The beneficial effects of this invention are as follows: This invention provides a target matching method for camera monitoring networks based on camera network topology. This method trains an existing target matching model, which does not have to be the most advanced, highest-performing, and largest-scale model. Instead, it can use a small model with good performance, thereby reducing the cost of application and making deployment more convenient and faster.

[0126] This invention improves matching accuracy by reordering the preliminary target matching image sequence output by the trained target matching model by utilizing information from the camera network topology, thereby balancing the performance and efficiency of the target matching method in complex situations.

[0127] This invention, based on the reordering of the feature distance metric formula of camera network topology, improves the accuracy of a single matching result while providing multiple accurate matching results, which is beneficial for mining target movement trajectories.

[0128] In the description of this invention, it should be understood that the terms "center," "thickness," "upper," "lower," "horizontal," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, a feature defined by "first," "second," and "third" may explicitly or implicitly include one or more of that feature.

Claims

1. A camera monitoring network target matching method based on a camera network topology, characterized in that, The method comprises the following steps: S1, collecting relevant scene data, constructing a training set according to target images in the relevant scene data, training a target matching network, and obtaining a target matching model; S2, inputting a target image to be matched into the target matching model to obtain a preliminary target matching image sequence; The S2 is specifically: inputting the target image to be matched and a plurality of search images into the target matching model to obtain a feature distance of the target image to be matched and each search image, and performing ascending sequence sorting on all search images according to the feature distance of the target image to be matched and each search image to obtain a preliminary target matching image sequence; The search images are obtained by searching the target image to be matched through a search database; In the S2, the target image to be matched p with the first i Zhang search image of the feature distance The expression is: In the formula, is a feature vector extracted from the target image to be matched by the target matching model, is a feature vector extracted from the first target image by the target matching model, i is a feature vector extracted from the Zhang search image by the target matching model, , N is a total number of search images, M is a semi-positive definite matrix, is a transpose operation; S3, reordering the preliminary target matching image sequence through a feature distance measurement formula based on a camera network topology to obtain a target matching result; The S3 comprises the following steps: S31, obtaining camera network distribution information, labeling cameras in the camera network distribution information according to the camera network distribution information, arranging each labeled camera according to the distance between the labeled cameras, obtaining a camera distance sequence corresponding to each labeled camera, and obtaining a distance sequence set of the cameras according to the camera distance sequences corresponding to all labeled cameras, which is specifically: with representing the camera acquiring the image of the target to be re-identified, representing the picture dataset in the retrieval database belonging to the camera of the first retrieval image, the total number of cameras, when the camera label of the target image to be matched is j , the camera label is arranged according to the distance between all cameras and the camera , and the distance sequence of the camera corresponding to the camera is obtained , wherein the closer the distance, the higher the ranking, that is, the current camera label is ranked first, and the distance sequence set of the camera is obtained ; S32, based on the distance sequence set of the cameras, obtaining a network topology weight of each labeled camera corresponding to the camera according to the camera distance sequence corresponding to each labeled camera, and obtaining a network topology weight set according to the network topology weights of all cameras, which is specifically: Based on the camera distance sequence corresponding to the camera j , get a camera network topology weight , that is, the weight of camera to camera is 1, the weight of camera to camera is 2, the weight of camera to camera is , get the network topology weight corresponding to each labeled camera in the camera network, get the network topology weight set ; S33, adding the network topology weight set to a K-neighbor-based reordering algorithm to obtain a feature distance measurement formula based on a camera network topology; In the S33, the characteristic distance measurement formula based on the camera network topology Specifically, the following formula: wherein cam is the target image to be matched, p is the camera label where the target image to be matched is located, is the camera, is the network topology weight value of the camera relative to the camera, , and , is the summation of the camera labels, a is the topology parameter, is the picture dataset of the search database, is the improved Jaccard distance, is the target image to be matched, p is the feature distance of the first search image i , and the second search image . The improved jaccard distance The expression is specifically: wherein is the target image to be matched p is the image in the search image database is the modified K-neighborhood image set value of the image in the search image database ranging between is the image in the search image database is the modified K-neighborhood image set value of the image in the search image database p ranging between is the image in the search image database ranging between S34, reordering the preliminary target matching image sequence according to the feature distance measurement formula based on the camera network topology to obtain a target matching result.

2. The camera surveillance network target matching method based on a camera network topology according to claim 1, characterized in that, The S1 comprises the following steps: S11, collecting relevant scene data, selecting target images in the relevant scene data, and using a random erasing method to randomly erase the target images to obtain a training set; S12, obtaining a target matching network, training the target matching network through the training set, and obtaining a target matching model.

3. The camera surveillance network target matching method based on camera network topology according to claim 2, characterized in that, In the S11, the random erasing method is specifically: SA1, presetting a random erasing probability parameter, selecting pictures to be randomly erased in the target images according to the random erasing probability parameter to obtain selected pictures; SA2, selecting a rectangular region in the selected pictures to obtain a selected rectangular region, filling pixel values of the selected rectangular region, and completing random erasing of the target images.

4. The camera surveillance network target matching method based on camera network topology according to claim 3, characterized in that, In the SA2, the method for obtaining the selected rectangular region is specifically: Select a point in the selected image. P , be on point P coordinates If the following formula is satisfied, then the point will be... P Build area As the selected rectangular area; wherein W is the length of the selected picture, H is the width of the selected picture, is the length of the selected rectangular region, is the width of the selected rectangular region; In the SA2, the method for filling the pixel values of the selected rectangular region is specifically: The value of each pixel point in the selected rectangular region is filled by filling pixel value, and the filling pixel value The expression is specifically: In the formula, is the value of each pixel point in the selected rectangular region, is the number of channels of the picture pixels, is the total number of pixel points of the selected rectangular region.

5. The camera surveillance network target matching method based on camera network topology according to claim 2, characterized in that, The S12 is specifically: obtaining a target matching network, training the target matching network through the training set, and adding camera network topology information to perform performance compensation to obtain a target matching model.

Images