Network site selection prediction method, device, equipment, medium and program product

Abstract

The present disclosure provides a method for predicting the location of an outlet, which can be applied to the fields of big data technology, target detection technology or finance. The method for predicting the location of an outlet includes: obtaining a target image containing a vehicle to be identified, wherein the target image includes m sub-images corresponding to the same geographical range at different time points, wherein m is an integer greater than or equal to 2; performing vehicle recognition on the target image based on a pre-trained vehicle recognition model to obtain a vehicle recognition result, wherein the vehicle recognition result includes the vehicle position, vehicle category and number of each type of vehicle corresponding to each sub-image; obtaining vehicle flow data based on the vehicle recognition result; and performing network location prediction based on the vehicle flow data, wherein the vehicle recognition model is obtained by training a YOLOv3 network improved by a deep residual module. The present disclosure also provides a network location prediction device, equipment, storage medium and program product.

Description

Technical Field

[0001] The present disclosure relates to the field of big data technology, target detection or finance, and specifically to a network site selection prediction method, device, equipment, medium and program product. Background Art

[0002] The location of outlets has a significant impact on the economic benefits, customer resource acquisition and market share of outlets. An important aspect of traditional outlet location selection is to conduct forecasts on the location of outlets by manually investigating the population flow and economic level of the target area. However, manual research is time-consuming, labor-intensive and costly, and manual statistical data is often incomplete and incomplete. Summary of the invention

[0003] In view of the above problems, the embodiments of the present disclosure provide a network site selection prediction method, apparatus, device, medium and program product for improving prediction intelligence and accuracy.

[0004] According to a first aspect of the present disclosure, a method for predicting network site selection is provided, comprising: obtaining a target image containing a vehicle to be identified, wherein the target image includes m sub-images corresponding to the same geographical range at different time points, wherein m is an integer greater than or equal to 2; performing vehicle recognition on the target image based on a pre-trained vehicle recognition model to obtain a vehicle recognition result, wherein the vehicle recognition result includes a vehicle position, a vehicle category, and a number of each type of vehicle corresponding to each sub-image; obtaining vehicle flow data based on the vehicle recognition result; and performing network site selection prediction based on the vehicle flow data, wherein the vehicle recognition model is obtained by training a YOLOv3 network improved based on a deep residual module.

[0005] According to an embodiment of the present disclosure, vehicle recognition is performed on the target image based on a pre-trained vehicle recognition model to obtain a vehicle recognition result, including: processing the target image based on a feature extraction network to obtain a first feature image, wherein the first feature image includes k groups of first sub-feature images of different target sizes, and k is an integer greater than or equal to 6; inputting the first image feature into a feature fusion network to obtain a second feature image; inputting the second feature image into a prediction layer network to obtain a predicted feature image; generating a priori target frame based on the predicted feature image; and processing the prior target frame based on a non-maximum suppression algorithm to obtain a vehicle recognition result.

[0006] According to an embodiment of the present disclosure, inputting the first feature image into a feature fusion network to obtain a second feature image includes: marking the kth group of first sub-feature maps obtained from the feature extraction network as the first group of second sub-feature maps; performing a 2-fold upsampling operation on the jth group of second sub-feature maps, and then fusing them with the kjth group of first sub-feature maps output by the feature extraction network in a cascade manner to obtain the sth group of second sub-feature maps, wherein j traverses integer values ​​in [1, k-1] in ascending order, s satisfies s=j+1 and s is an integer; and obtaining a second feature image based on the 1st to kth groups of second sub-feature maps.

[0007] According to an embodiment of the present disclosure, the feature extraction network includes a residual module and a convolutional layer, wherein the number of residual module groups is n, n satisfies k≤n and n is an integer, the number of convolutional layers is associated with the number of residual module groups, and the convolutional layers include convolutional layers that satisfy the target field of view.

[0008] According to an embodiment of the present disclosure, the target image is processed based on the feature extraction network to obtain the first feature image, including: inputting the target image into the feature extraction network, and obtaining a first group of output images after processing by a convolution layer; processing the first group of output images based on a residual module to obtain a second group of output images; alternately executing the convolution layer processing steps and the residual module processing steps until the nth group of output images is obtained; and obtaining the first feature image based on the second to nth groups of output images.

[0009] According to an embodiment of the present disclosure, processing the target image based on a feature extraction network to obtain a first feature image further includes: discarding output images that do not meet the target size.

[0010] According to an embodiment of the present disclosure, obtaining the first feature image based on the second to nth groups of output images further includes: discarding the second to nkth groups of output images to obtain k groups of first sub-feature images of different target sizes.

[0011] According to an embodiment of the present disclosure, before performing vehicle identification, the method also includes a step of preprocessing the target image, and the preprocessing includes adjusting the size of the target image so that the adjusted target image can output a first feature image of the target size after being processed by a feature extraction network.

[0012] According to an embodiment of the present disclosure, the first feature image includes 6 groups of first sub-feature maps, and the sizes of the first sub-feature maps are 64×64, 32×32, 16×16, 8×8, 4×4, and 2×2, respectively.

[0013] According to an embodiment of the present disclosure, the size of the target image is 512×512.

[0014] According to an embodiment of the present disclosure, the convolution layer does not include a 1×1 convolution layer.

[0015] According to an embodiment of the present disclosure, the geographical range includes q target areas, q is an integer greater than or equal to 2, and a mapping relationship exists between the vehicle position and the target area. The method of obtaining traffic flow data based on the vehicle recognition result also includes: counting the number of each type of vehicles in the pth target area in the sub-image at the i-th time point, marking it as first traffic flow data, wherein p satisfies 1≤p≤q, and i satisfies 1≤i≤m; obtaining second traffic flow data based on the first traffic flow data of the sub-images at m different time points, wherein the second traffic flow data includes the first traffic flow data of the sub-image at each time point; processing the second traffic flow data based on a preset algorithm to obtain the traffic flow data, wherein the traffic flow data and the target area have a mapping relationship.

[0016] According to an embodiment of the present disclosure, the vehicle flow data at least includes automobile flow data.

[0017] According to an embodiment of the present disclosure, the network point location prediction based on the vehicle flow data includes: sorting the vehicle flow data corresponding to q target areas in order from large to small; and obtaining the network point location prediction result based on the sorting result.

[0018] The second aspect of the present disclosure provides a network site selection prediction device, including: an acquisition module, configured to acquire a target image containing a vehicle to be identified, wherein the target image includes m sub-images corresponding to the same geographical range at different time points, wherein m is an integer greater than or equal to 2; an identification module, configured to perform vehicle identification on the target image based on a pre-trained vehicle identification model to obtain a vehicle identification result, wherein the vehicle identification result includes a vehicle position, a vehicle category, and a number of each type of vehicle corresponding to each sub-image, and the vehicle identification model is obtained by training a YOLOv3 network improved by a deep residual module; a first processing module, configured to acquire vehicle flow data based on the vehicle identification result; and a second processing module, configured to perform network site selection prediction based on the vehicle flow data.

[0019] The third aspect of the present disclosure provides an electronic device, comprising: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the above-mentioned network site selection prediction method.

[0020] The fourth aspect of the present disclosure further provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, causes the processor to execute the above-mentioned network site selection prediction method.

[0021] The fifth aspect of the present disclosure also provides a computer program product, including a computer program, which implements the above-mentioned network site selection prediction method when executed by a processor.

[0022] The method provided by the embodiments of the present invention takes vehicle flow data as an important indicator for network site selection prediction, and obtains vehicle flow data of the target area by using a vehicle recognition model obtained by training the YOLOv3 network improved by a deep residual module. This reduces the manpower and time costs of traditional manual survey methods to a certain extent, and improves the intelligence and accuracy of network site prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above contents and other objects, features and advantages of the present disclosure will become more apparent through the following description of the embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0024] Figure 1 The application scenario diagram of the network site selection prediction method, device, equipment, medium and program product according to the embodiments of the present disclosure is schematically shown.

[0025] Figure 2 The flowchart of the network point location prediction method according to the embodiment of the present disclosure is schematically shown.

[0026] Figure 3 The flowchart schematically shows a method for performing vehicle recognition on the target image based on the pre-trained vehicle recognition model to obtain a vehicle recognition result according to an embodiment of the present disclosure.

[0027] Figure 4 The flowchart of the method for processing a target image based on a feature extraction network and obtaining a first feature image according to an embodiment of the present disclosure is schematically shown.

[0028] Figure 5 A flowchart of a method for acquiring a first feature image based on the second to nth groups of output images according to some specific embodiments of the present disclosure is schematically shown.

[0029] Figure 6 The flowchart of the method of inputting the first feature image into the feature fusion network to obtain the second feature image according to an embodiment of the present disclosure is schematically shown.

[0030] Figure 7 The flowchart of the method for preprocessing a target image according to an embodiment of the present disclosure is schematically shown.

[0031] Figure 8 A feature extraction network of an example of the present disclosure is schematically shown.

[0032] Fig. 9 The network structure diagram of the vehicle identification model of the example disclosed in the present invention is schematically shown.

[0033] Fig.10 A flowchart of a method for acquiring vehicle flow data based on vehicle recognition results according to an embodiment of the present disclosure is schematically shown.

[0034] Fig.11 A flowchart of a method for predicting network site selection based on the vehicle flow data according to an embodiment of the present disclosure is schematically shown.

[0035] Fig.12 The structural block diagram of the network point location prediction device according to an embodiment of the present disclosure is schematically shown.

[0036] Fig.13 A block diagram of an electronic device suitable for implementing a network site selection prediction method according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0037] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0038] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0039] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0040] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0041] The location of bank branches has a significant impact on the economic benefits of the branches, customer resource acquisition and market share. In the process of site selection evaluation, the population flow and economic level of the target area are important factors affecting the decision. The traditional method of branch site selection is to manually survey the population flow and economic level of the target area. However, manual surveys are time-consuming, labor-intensive, and costly, and manual statistical data are often incomplete and incomplete.

[0042] In the process of studying and obtaining the embodiments of the present disclosure, the inventors found that the traffic volume around the outlets is related to the population flow and economic level of the outlet area. The more traffic volume, especially the more car traffic, the better the location of the outlet, which determines the number of customers of the outlet to a certain extent. Therefore, the more traffic volume can provide valuable information for the site selection of bank outlets. On this basis, the site selection of outlets can be predicted based on the traffic volume data, which can solve the problem that traditional research methods are difficult to obtain comprehensive, scientific and accurate data to a certain extent. Traditional traffic volume data is often obtained by human research methods, which is time-consuming and labor-intensive. With the development of deep learning, some models that can better identify vehicle categories have been increasingly used. Using vehicle recognition results to conduct traffic volume data research will greatly reduce labor costs and improve the accuracy and efficiency of research results.

[0043] Among the vehicle recognition models based on deep learning, the more classic models include Faster R-CNN and YOLOv3. The YOLOv3 model has greater advantages than Faster R-CNN in terms of detection accuracy and detection speed. However, YOLOv3 still has the problem of low detection accuracy for small-sized vehicle targets such as cars.

[0044] In view of the above-mentioned problems existing in the prior art, an embodiment of the present invention provides a network site selection prediction method, comprising: obtaining a target image containing a vehicle to be identified, wherein the target image includes m sub-images corresponding to the same geographical range at different time points, wherein m is an integer greater than or equal to 2; performing vehicle recognition on the target image based on a pre-trained vehicle recognition model to obtain a vehicle recognition result, wherein the vehicle recognition result includes the vehicle position, vehicle category and number of each type of vehicle corresponding to each sub-image; obtaining vehicle flow data based on the vehicle recognition result; and performing network site selection prediction based on the vehicle flow data, wherein the vehicle recognition model is obtained by training a YOLOv3 network improved based on a deep residual module.

[0045] It should be noted that the network site selection prediction method, device, equipment, medium and program product provided in the embodiments of the present disclosure can be used in big data technology, target detection technology in the acquisition of vehicle flow data, and can also be used in a variety of fields other than big data technology and target detection technology, such as the financial field, especially network site selection prediction, etc. The application field of the network site selection prediction method, device, equipment, medium and program product provided in the embodiments of the present disclosure is not limited.

[0046] The above operations for achieving at least one purpose of the present disclosure will be described below in conjunction with the accompanying drawings and their explanatory texts.

[0047] Figure 1 The application scenario diagram of the network site selection prediction method, device, equipment, medium and program product according to the embodiments of the present disclosure is schematically shown.

[0048] like Figure 1 As shown, the application scenario 100 according to this embodiment may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0049] The user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, etc. The terminal devices 101, 102, 103 may have an image acquisition function. The terminal devices 101, 102, 103 may also have various communication client applications, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0050] The terminal devices 101, 102, 103 may be various electronic devices with image acquisition functions, including but not limited to cameras, webcams, smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0051] The server 105 may be a server that provides various services, such as a background management server (only as an example) that provides support for websites browsed by users using the terminal devices 101, 102, and 103. The background management server may analyze and process the received data such as user requests, and may also feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal device.

[0052] It should be noted that the network site selection prediction method provided in the embodiment of the present disclosure can generally be executed by the server 105. Accordingly, the network site selection prediction device provided in the embodiment of the present disclosure can generally be set in the server 105. The network site selection prediction method provided in the embodiment of the present disclosure can also be executed by a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105. Accordingly, the network site selection prediction device provided in the embodiment of the present disclosure can also be set in a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105.

[0053] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to the implementation requirements.

[0054] The following will be based on Figure 1 The scene described by Figure 2 to Figure 11 The network site selection prediction method of the disclosed embodiment is described in detail.

[0055] Figure 2 The flowchart of the network point location prediction method according to the embodiment of the present disclosure is schematically shown.

[0056] like Figure 2 As shown, the network point location prediction method of this embodiment includes operations S210 to S240. The network point location prediction method can be executed by a processor or any electronic device including a processor.

[0057] In operation S210 , a target image containing a vehicle to be recognized is acquired.

[0058] According to an embodiment of the present disclosure, the target image includes m sub-images corresponding to the same geographical range at different time points, where m is an integer greater than or equal to 2.

[0059] In operation S220, vehicle recognition is performed on the target image based on the pre-trained vehicle recognition model to obtain a vehicle recognition result.

[0060] According to an embodiment of the present disclosure, the vehicle recognition result includes the vehicle position, vehicle category and the number of each type of vehicle corresponding to each sub-image.

[0061] In operation S230, vehicle flow data is acquired based on the vehicle recognition result.

[0062] In operation S240, network point location prediction is performed based on the vehicle flow data.

[0063] In an embodiment of the present disclosure, the target image may be an image of the same geographical range at different time points obtained according to a preset observation period. The geographical range may be an area to be investigated, such as a city, a street, an administrative area, etc. The target image may be obtained from a vehicle data set of the area to be investigated, which may be obtained based on surveillance cameras, drone aerial photography, satellite remote sensing images, etc. In an embodiment of the present disclosure, the target image may mainly be a picture of traffic flow in various areas of the city taken by drone aerial photography. It can be understood that before vehicle identification, the vehicle data set image may be preprocessed to filter out clear images containing vehicle targets as the target image. The setting of the observation period may improve the accuracy of the acquired traffic flow data to achieve a more accurate assessment of the site selection of the network, which may be set based on expert experience. The vehicle identification model of the embodiment of the present disclosure may be obtained by training a YOLOv3 network improved by a deep residual module. In the features of model identification, a deeper network layer may reduce the error of the network. However, as the number of network layers increases, network degradation may occur. The residual structure can effectively solve this problem. The shortcut link used by it can strengthen the transmission of features and reduce information loss, thereby increasing the feature reuse of the feature extraction network and providing richer target features for the subsequent detection network. The embodiment of the present disclosure draws on the residual idea of ​​ResNet and improves the feature extraction network in YOLOv3. Specifically, a feature extraction network with a deep residual network as the main improvement method is constructed in YOLOv3. More specifically, some residual modules and more convolutional layers are added to DarkNet-53 to extract the features of vehicle targets, thereby achieving more accurate vehicle recognition, especially improving the recognition ability of small-sized targets such as cars. Among them, in the process of training the vehicle recognition model of the embodiment of the present disclosure, after preprocessing the vehicle data set to obtain the target image, a certain number of target images (e.g., 80%) can be selected as the training set, and the remaining target images (e.g., 20%) can be used as the test set. Use the training set to train the model. When the loss function value tends to be stable, stop training and obtain the best prediction model as the final vehicle recognition model. The vehicle recognition model of the embodiment of the present disclosure can mark the vehicle position, vehicle category and the number of each category of vehicles corresponding to each sub-image. Among them, the vehicle position can be obtained based on the predicted target box of the vehicle recognition model. In this way, the vehicle distribution status of each sub-image can be obtained, and further, the size of the traffic volume of any target area in the area to be investigated can be determined. Furthermore, the key target area suitable for network site selection can be predicted. It can be understood that after the observation period, each target area can be further evaluated based on the target image within the period, so as to facilitate the prediction of the area most suitable as the network site selection location.

[0064] Figure 3The flowchart schematically shows a method for performing vehicle recognition on the target image based on the pre-trained vehicle recognition model to obtain a vehicle recognition result according to an embodiment of the present disclosure.

[0065] like Figure 3 As shown, the method for obtaining a vehicle recognition result in this embodiment includes operations S310 to S350.

[0066] In operation S310, the target image is processed based on a feature extraction network to obtain a first feature image, wherein the first feature image includes k groups of first sub-feature images of different target sizes, where k is an integer greater than or equal to 6.

[0067] In operation S320, the first image features are input into a feature fusion network to obtain a second feature image.

[0068] In operation S330, the second feature image is input into a prediction layer network to obtain a predicted feature image.

[0069] In operation S340 , a priori target box is generated based on the predicted feature image.

[0070] In operation S350, the priori target frame is processed based on a non-maximum suppression algorithm to obtain a vehicle recognition result.

[0071] According to an embodiment of the present disclosure, the YOLOv3 network improved based on the deep residual module includes a feature extraction network, a feature fusion network and a prediction layer network. The feature extraction network of the traditional YOLOv3 network is a DarkNet-53 network. The embodiment of the present disclosure adds a residual module and more convolutional layers to the feature extraction network. Among them, the shortcut link used by the residual module can strengthen the transmission of features and reduce information loss, thereby increasing the feature reuse of the feature extraction network and providing rich target features for the subsequent detection network. As a result, more residual modules can provide richer target features. With the increase of residual modules in the feature extraction network, the feature fusion network can fuse more sizes of first sub-feature maps to output more sizes of second feature images. Multi-scale feature fusion enables the feature information of feature maps at different levels to be fully integrated and utilized. Each group of feature maps of the size of the second feature image is a group of detection branches, and each group of detection branches can independently perform vehicle recognition tasks using contextual semantic information, and each group of branches shares the features extracted from the feature extraction network, enhancing the expression ability of network features. The prediction layer network includes a convolutional layer. After convolution processing is performed on the second feature image, a predicted feature image is finally obtained for vehicle identification. When making a prediction, the predicted feature image can be divided into N×N grids according to the size (the scale of the predicted feature image is different, and the size of N is also different), and a preset number of different prior frames are configured for each grid. Compared with the traditional YOLOv3 network, the improved YOLOv3 network provided in the embodiment of the present disclosure can generate more prior frames, which improves the detection effect of vehicle targets to a certain extent. The vehicle position of the embodiment of the present disclosure is calibrated by the vehicle target frame. Among them, the vehicle target frame is generated by processing the prior target frame using the non-maximum suppression algorithm. The non-maximum suppression algorithm suppresses redundant prior target frames through an iteration-traversal-elimination process. The specific process mainly includes: sorting the scores of all prior target frames, selecting the highest score and its corresponding frame. Traverse the remaining frames, and if the overlapping area with the current highest score frame is greater than the preset threshold, the frame is deleted. Continue to select the highest score from the unprocessed frame, and repeat the above comparison-deletion process. Therefore, after removing duplicate or erroneous bounding boxes using the non-maximum suppression method, a bounding box that retains the correct classification and location information can be obtained, thereby improving the accuracy of the prior target box annotation and obtaining the final vehicle recognition result.

[0072] The feature extraction network designed in the embodiment of the present disclosure includes a residual module and a convolution layer, wherein the number of residual module groups is n, n satisfies k≤n and n is an integer. Thus, after obtaining the feature recognition result based on the residual module, the recognition result can be screened to meet the vehicle recognition requirements of the embodiment of the present disclosure. Among them, the number of layers of the convolution layer is associated with the number of residual module groups. For example, after being processed by one or more convolution layers, the first group of residual modules is connected for processing. Further, a group of convolution layer processing and a group of residual module processing are alternately completed to output feature maps of different sizes. Thus, the number of layers a of the convolution layer can satisfy a>n and a is an integer, for example, a=n+1. In the embodiment of the present disclosure, the convolution layer may include a convolution layer that satisfies the target field of view to obtain a feature map that satisfies the target size. It is worth noting that at this time, the convolution layer that does not meet the target field of view can be discarded to simplify the network structure, reduce data processing overhead, and improve vehicle recognition efficiency. Among them, the convolution layer that does not meet the target field of view can be a convolution layer that has little effect on identifying the vehicle of the embodiment of the present disclosure.

[0073] In some specific embodiments, the convolution layer does not include a 1×1 convolution layer. The feature extraction network of the embodiment of the present disclosure can output a feature map with a minimum size of 1×1, and the area mapped on the input image is close to the size of the entire image. However, there are no vehicle targets close to the size of the entire image in the vehicle data set to be detected, which deviates from the actual needs. Therefore, the feature extraction network of the embodiment of the present disclosure abandons the 1×1 convolution layer, which has little effect on the final detection accuracy of the model. At the same time, because the computational complexity is reduced, the recognition speed of the model will be improved to a certain extent.

[0074] Figure 4 The flowchart of the method for processing a target image based on a feature extraction network and obtaining a first feature image according to an embodiment of the present disclosure is schematically shown.

[0075] like Figure 4 As shown, the method of acquiring a first feature image by processing a target image based on a feature extraction network in this embodiment includes operations S410 to S440.

[0076] In operation S410, the target image is input into the feature extraction network, and a first set of output images is obtained after being processed by a convolutional layer.

[0077] In operation S420, the first group of output images is processed based on a residual module to obtain a second group of output images.

[0078] In operation S430, the convolution layer processing step and the residual module processing step are performed alternately until the nth group of output images is obtained.

[0079] In operation S440, a first feature image is acquired based on the second to nth groups of output images.

[0080] According to an embodiment of the present disclosure, the feature extraction method is performed alternately by a convolution layer and a residual module. For example, features can be first extracted by one or more convolution layers. For example, after filtering by a filter of a first size in the first convolution layer, the result is input to the next convolution layer, and the second convolution layer uses a filter of a second size to filter the input data, and then inputs the next convolution layer, until the convolution layer of the previous layer of the first group of residual modules. After the filter in the convolution layer of the previous layer of the first group of residual modules filters the input data of this layer and performs a 2-fold downsampling operation, a first group of output images is obtained. The first group of output images is input into the first group of residual modules for processing to obtain a second group of output images. The residual module used in the embodiment of the present disclosure is similar to the residual module used in YOLOv3, which includes structures such as a convolution layer, a BN layer, and an activation function LeakyReLU, and is connected in a residual manner. After obtaining the first group of output images. Continue to process n-1 times by alternating processing by a convolution layer and a group of residual modules until the nth group of output images is obtained. Among them, the processing flow of each group of convolution layer processing + residual module processing is to perform convolution layer filtering, filter it and perform a 2-fold downsampling operation before inputting it into the residual module processing. The number of residual modules in each group of residual modules is preset. Except for the different filter and output image sizes, the other structures of these residual modules are similar, thereby obtaining the second to n-th groups of output images. Further, the first feature image can be obtained based on the second to n-th groups of output images. For example, the second to n-th groups of output images together constitute the first feature image.

[0081] According to some specific embodiments, processing the target image based on a feature extraction network to obtain a first feature image further includes: discarding output images that do not meet the target size.

[0082] According to a specific embodiment of the present disclosure, in the second to nth groups of output images, some shallow output images have a small receptive field and have no practical value for detecting large and medium-sized vehicle sizes in the vehicle data set. Therefore, output images that do not meet the target size can be discarded, for example, a preset number of shallow output images, so as to reduce data processing overhead while meeting the requirements for identifying vehicles of various specifications. It should be understood that the discarded output images that do not meet the target size are one or more groups of the second to nth groups of output images.

[0083] Figure 5 A flowchart of a method for acquiring a first feature image based on the second to nth groups of output images according to some specific embodiments of the present disclosure is schematically shown.

[0084] like Figure 5As shown, the method for acquiring the first feature image based on the second to nth groups of output images in this specific embodiment includes operation S510.

[0085] In operation S510, the second to nkth groups of output images are discarded to obtain k groups of first sub-feature maps of different target sizes. It can be understood that when n>k, the second to nkth groups of output images can be discarded as shallow output images of little use to reduce data processing overhead and improve data processing efficiency.

[0086] In a specific example, the first feature image includes 6 groups of first sub-feature maps, and the sizes of the first sub-feature maps are 64×64 pixels, 32×32 pixels, 16×16 pixels, 8×8 pixels, 4×4 pixels, and 2×2 pixels, respectively. The sizes of the above 6 groups of sub-feature maps can basically cover vehicle targets of different sizes in the vehicle data set, and fully reflect the feature information of large, medium and small vehicle targets in the vehicle data set, so that the feature extraction network can better complete the feature extraction task.

[0087] Figure 6 The flowchart of the method of inputting the first feature image into the feature fusion network to obtain the second feature image according to an embodiment of the present disclosure is schematically shown.

[0088] like Figure 6 As shown, the method for acquiring the second characteristic image in this embodiment includes operations S610 to S630.

[0089] In operation S610, the kth group of first sub-feature maps obtained from the feature extraction network is marked as the first group of second sub-feature maps.

[0090] In operation S620, after performing a 2-fold upsampling operation on the j-th group of second sub-feature maps, the maps are fused in a cascade manner with the kj-th group of first sub-feature maps output by the feature extraction network to obtain the s-th group of second sub-feature maps, where j traverses integer values ​​in [1, k-1] in ascending order, and s satisfies s=j+1 and is an integer.

[0091] In operation S630, a second feature image is acquired based on the 1st to kth groups of second sub-feature images.

[0092] According to an embodiment of the present disclosure, a multi-scale feature fusion strategy similar to that in YOLOv3 is used to fuse the first feature image from the feature extraction network to make full use of the feature information of feature maps at different levels. Since the feature extraction network outputs more sub-feature maps with more matching sizes, the fused second feature image also has better prediction capabilities for the application scenarios of the present disclosure. In the fusion process, the last group of sub-feature maps is used as the first group of second sub-feature maps, which can also be used as a feature image for prediction. Further, starting from the last group of first sub-feature maps, each group of first sub-feature maps is subjected to a 2-fold upsampling operation and then fused with the previous group of first sub-feature maps in a cascade manner to obtain the fused second sub-feature map. After repeating the above fusion process k-1 times, k groups of second sub-feature maps can be obtained. Specifically, the kth group of first sub-feature maps is used as the first group of second sub-feature maps, and after performing a 2-fold upsampling operation, it is cascaded and fused with the k-1th group of first sub-feature maps to obtain the second group of second sub-feature maps. After performing a 2-fold upsampling operation on the second group of second sub-feature maps, they are cascaded and fused with the k-2th group of first sub-feature maps to obtain the third group of second sub-feature maps. After performing a 2-fold upsampling operation on the third group of second sub-feature maps, they are cascaded and fused with the k-3th group of first sub-feature maps to obtain the fourth group of second sub-feature maps. And so on, until the k-1th group of second sub-feature maps are performed a 2-fold upsampling operation, they are cascaded and fused with the first group of first sub-feature maps to obtain the kth group of second sub-feature maps. Since starting from the second group of second sub-feature maps, each group of second sub-feature maps utilizes the information of the previous groups of feature maps, and the k groups of second sub-feature maps share the features extracted from the feature extraction network, the expression ability of the network features is enhanced. It should be understood that each group of sub-feature maps in the second feature image can independently perform the vehicle recognition task using contextual semantic information, thereby identifying various types of vehicles of different sizes and specifications.

[0093] According to an embodiment of the present disclosure, before performing vehicle recognition, the method further includes a step of preprocessing the target image.

[0094] Figure 7 The flowchart of the method for preprocessing a target image according to an embodiment of the present disclosure is schematically shown.

[0095] like Figure 7 As shown, the method for preprocessing a target image in this embodiment includes operation S710.

[0096] In operation S710, the size of the target image is adjusted so that the adjusted target image can output a first feature image of a target size after being processed by a feature extraction network.

[0097] In the embodiment of the present disclosure, the target image is used as an input image, and its size is associated with the number and size of the first feature images that can be obtained. In order to output the first feature image of the target size after being processed by the feature extraction network, the size of the target image can be adjusted. Among them, after increasing the size of the target image, it can be input into the feature extraction network improved by the deep residual module, and more first feature images that meet the target size requirements can be output to complete the vehicle recognition task of the embodiment of the present disclosure.

[0098] In a specific example, the size of the target image is 512×512 pixels. Thus, nine first feature images of different sizes can be obtained to meet the accurate recognition of various vehicle types and specifications such as cars, trucks, buses, and trucks.

[0099] Combine the following Figure 8-Figure 9 An exemplary detailed description of an exemplary method of the present disclosure is to perform vehicle recognition on the target image based on a pre-trained vehicle recognition model to obtain a vehicle recognition result. It should be understood that the following description is only an exemplary description and does not constitute a specific limitation of the present disclosure.

[0100] In an example of the present disclosure, in order to unify the size of the fused feature map so that the feature extraction network can output feature maps of more sizes to meet the needs of vehicle detection in this example, the size of the input image of size 416×416 used in the original YOLOv3 is adjusted to 512×512 as the target image size of the example of the present disclosure. The above adjustment enables the feature extraction network used in this example to output feature maps of 10 sizes (1×1 pixels, 2×2 pixels, 4×4 pixels, 8×8 pixels, 16×16 pixels, 32×32 pixels, 64×64 pixels, 128×128 pixels, 256×256 pixels, 512×512 pixels), which has better performance in vehicle recognition than the feature maps of 13×13 pixels, 26×26 pixels, 52×52 pixels, 104×104 pixels, 208×208 pixels, and 416×416 pixels that can be extracted by the traditional YOLOv network. Among them, the feature map with the minimum size of 1×1 has the largest receptive field, which is conducive to the detection of large-sized targets. However, in the example disclosed in this disclosure, since the minimum feature map size is 1×1, the area mapped on the input image is close to the size of the entire image, and there are no vehicle targets close to the size of the entire image in the vehicle dataset, which deviates from the actual needs. Therefore, the feature extraction network designed in this example abandons the 1×1 convolution layer, which has little effect on the final detection accuracy of the model. At the same time, because the computational complexity is reduced, the detection speed of the model will be improved to a certain extent.

[0101] Figure 8 A feature extraction network of an example of the present disclosure is schematically shown.

[0102] like Figure 8 As shown in the figure, the target image size is 512×512. The first convolutional layer uses 16 3×3 filters to filter the input image, and then outputs the result to the next convolutional layer. The second convolutional layer uses 32 3×3 filters and performs a downsampling operation with a step size of 2. The result is input into the first group of residual modules (including one residual module). This residual module is similar to the residual module used in YOLOv3, which contains structures such as convolutional layers, BN layers, and activation functions Leaky ReLU, and is connected in a residual manner. The feature map size obtained after a residual module is 256×256. Next, continue to execute 7 groups of networks with 2, 8, 8, 4, 4, 4, and 2 residual modules respectively. These residual modules are similar except for the different sizes of filters and output feature maps. In addition, each group of residual modules is preceded by a downsampling operation with a step size of 2, so that 8 groups of feature maps of different sizes can be obtained.

[0103] Fig. 9 The network structure diagram of the vehicle identification model of the example disclosed in the present invention is schematically shown.

[0104] like Fig. 9 As shown, the network structure of the vehicle recognition model of the example disclosed in the present invention includes a feature extraction network, a feature fusion network and a prediction layer network. Figure 8 Corresponding. Fig. 9In , CBL represents convolution block, Res represents residual module, and the number of residual components contained in this group of residual modules is represented by Conv. Conv represents convolution layer. In the example disclosed in the present invention, cascade fusion is used for fusion. As an optimization of the example disclosed in the present invention, the receptive field of the shallow feature map is too small, and it has no practical value for identifying vehicles of different sizes in the vehicle data set. Therefore, this example discards the three groups of feature maps output by the first three groups of residual modules and only selects the following six groups of feature maps. The sizes of each group of feature maps are 64×64, 32×32, 16×16, 8×8, 4×4, and 2×2, respectively. These six groups of feature map sizes can basically cover vehicle targets of different sizes in the target image, so as to fully reflect the feature information of large, medium and small vehicle targets, so that the feature extraction network can better complete the feature extraction task and realize accurate classification of vehicles of various types and sizes. After obtaining the six groups of feature maps from the feature extraction network of the example disclosed in the present invention, the fusion of feature maps of six different scales can be realized to fully integrate and utilize the feature information of feature maps of different levels. Specifically, for feature maps of sizes 2×2, 4×4, 8×8, 16×16, and 32×32, a 2-fold upsampling operation is performed in sequence, and then feature fusion is performed with feature maps of sizes 4×4, 8×8, 16×16, 32×32, and 64×64 extracted by the feature extraction network, respectively, to obtain 6 sets of feature map detection branches, whose sizes are 2×2, 4×4, 8×8, 16×16, 32×32, and 64×64, respectively. Each set of branches can independently perform the vehicle target detection task using contextual semantic information, and these 6 sets of branches share the features extracted from the feature extraction network to enhance the network's expression ability. After the prediction layer network performs convolution processing on the 6 new feature map detection branches, the final predicted feature image can be obtained to achieve vehicle recognition.

[0105] In the vehicle recognition stage, the predicted feature image is divided into N×N grids. Among them, the size of N is different for different scales of the predicted feature image. In this example, it can be divided into 2×2 grids, 4×4 grids, 8×8 grids, 16×16 grids, 32×32 grids and 64×64 grids based on the size of the predicted feature image. Three different prior target boxes are configured for each grid. Thus, for each target image, 6 feature maps will generate 16,380 prior target boxes, which is more than the 10,647 prior target boxes generated by the traditional YOLOv3 at most. More prior target boxes can improve the detection effect of vehicle targets to a certain extent. Finally, the non-maximum suppression method is used to remove repeated or erroneous prior target boxes, and the prior target box that retains the correct classification and location information is obtained as the vehicle bounding box, and the final vehicle recognition result is obtained. In the example disclosed in the present invention, the vehicle bounding box can be used to identify the vehicle location. At the same time, the vehicle recognition result also includes the vehicle category and the number of each type of vehicle. The vehicle categories of the disclosed examples mainly include cars, vans, buses, and trucks, in order to collect statistics on traffic flow data suitable for network site selection prediction.

[0106] To verify the effectiveness of the vehicle recognition model in this example, this example uses an aerial vehicle traffic dataset as the experimental dataset, uses the four vehicle categories (cars, vans, buses, and trucks) in the dataset as detection targets, and uses recognition speed and average recognition accuracy as evaluation indicators to compare and analyze the recognition effects of the two methods.

[0107] Table 1 shows the comparison of the detection results of the vehicle recognition model of this example and the traditional YOLOv3 recognition model.

[0108] Model Recognition speed (fps) Mean average recognition accuracy (mAP (%)) YOLOv3 47.7 88.67 This example 44.5 93.47

[0109] Table 1

[0110] As shown in Table 1, the example vehicle recognition model of the present disclosure can achieve faster recognition speed and higher average recognition accuracy than the traditional YOLOv3 network while ensuring real-time recognition, thereby better predicting traffic flow and providing an important reference for network site selection.

[0111] According to an embodiment of the present disclosure, after obtaining the vehicle recognition result, the traffic flow data can be further obtained based on the vehicle recognition result. Wherein, the target image comes from the same geographical range. The geographical range includes q target areas, q is an integer greater than or equal to 2, and there is a mapping relationship between the vehicle position and the target area. Therefore, by identifying the vehicle in the target area and obtaining the traffic flow data of the target area, the traffic flow data of different target areas can be compared to predict the site selection of the network point.

[0112] Fig.10A flowchart of a method for acquiring vehicle flow data based on vehicle recognition results according to an embodiment of the present disclosure is schematically shown.

[0113] like Fig.10 As shown, the method for acquiring vehicle flow data based on vehicle recognition results in this embodiment includes operations S1010 to S1030.

[0114] In operation S1010, the number of vehicles of each type in the p-th target area in the sub-image at the i-th time point is counted and marked as first vehicle flow data.

[0115] In operation S1020, second traffic flow data is acquired based on the first traffic flow data of the m sub-images at different time points.

[0116] In operation S1030, the second vehicle flow data is processed based on a preset algorithm to obtain the vehicle flow data.

[0117] According to an embodiment of the present disclosure, the i-th time point is an element in the m time points, and the p-th target area is an element in the q target areas. It should be understood that p satisfies 1≤p≤q, and i satisfies 1≤i≤m. Since the vehicle recognition result contains a vehicle target frame that can mark the vehicle position, that is, the vehicle position can form a corresponding mapping relationship with the vehicle target frame. Thus, the number of each type of vehicle in each target area in the sub-image at the i-th time point can be obtained, and it can be marked as the first traffic data. Further, the number of each type of vehicle in each target area at each time point can be obtained. The first traffic data at each time point can be summarized to obtain the second traffic data. After processing the second traffic data, the traffic flow data can be obtained. It can be understood that there is a mapping relationship between the traffic flow data and the target area. Thus, the traffic flow of each target area can be obtained, so that the network site selection prediction can be performed based on the traffic flow. It should be noted that the preset algorithm can be an algorithm in the prior art, which is based on the ability to obtain traffic flow data, and the embodiment of the present disclosure does not impose any special restrictions on this.

[0118] In the embodiments of the present disclosure, the traffic flow data at least includes automobile traffic flow data. In some specific embodiments, when using the traffic flow data to predict the site selection of outlets, the automobile traffic flow data can be mainly screened. Combined with the goals and needs of the outlet business, the larger the traffic flow, especially the larger the automobile traffic flow, the more potential outlet customers there should be, which can provide a more important reference for the site selection of outlets.

[0119] Fig.11 A flowchart of a method for predicting network site selection based on the vehicle flow data according to an embodiment of the present disclosure is schematically shown.

[0120] like Fig.11As shown, the method for predicting network site selection based on the vehicle flow data in this embodiment includes operations S1110 to S1120.

[0121] In operation S1110 , the traffic flow data corresponding to the q target areas are sorted in descending order.

[0122] In operation S1120, a network site selection prediction result is obtained based on the sorting result.

[0123] According to an embodiment of the present disclosure, after obtaining the traffic flow data of all q target areas, the traffic flow can be sorted. Areas with large traffic flow should have more potential outlet customers. Based on the sorting results, a preset number of top-ranked target areas can be used as potential outlet site selection prediction results. The sorting results can also be processed and screened based on business needs to obtain the final outlet site selection prediction results.

[0124] Based on the above-mentioned network site selection prediction method, the present disclosure also provides a network site selection prediction device. Fig.12 The device is described in detail.

[0125] Fig.12 The structural block diagram of the network point location prediction device according to an embodiment of the present disclosure is schematically shown.

[0126] like Fig.12 As shown, the network site selection prediction device 800 of this embodiment includes an acquisition module 810 , an identification module 820 , a first processing module 830 and a second processing module 840 .

[0127] The acquisition module 810 is configured to acquire a target image containing a vehicle to be identified, wherein the target image includes m sub-images at different time points corresponding to the same geographical range, wherein m is an integer greater than or equal to 2.

[0128] The recognition module 820 is configured to perform vehicle recognition on the target image based on a pre-trained vehicle recognition model to obtain a vehicle recognition result, wherein the vehicle recognition result includes the vehicle position, vehicle category and number of each type of vehicle corresponding to each sub-image, and the vehicle recognition model is obtained by training a YOLOv3 network improved by a deep residual module.

[0129] The first processing module 830 is configured to obtain vehicle flow data based on the vehicle recognition result.

[0130] The second processing module 840 is configured to perform network site selection prediction based on the vehicle flow data.

[0131] According to an embodiment of the present disclosure, any multiple modules among the acquisition module 810, the identification module 820, the first processing module 830 and the second processing module 840 can be combined in one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present disclosure, at least one of the acquisition module 810, the identification module 820, the first processing module 830 and the second processing module 840 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware and firmware or in a suitable combination of any of them. Alternatively, at least one of the acquisition module 810, the identification module 820, the first processing module 830 and the second processing module 840 may be at least partially implemented as a computer program module, and when the computer program module is executed, a corresponding function may be performed.

[0132] Fig.13 A block diagram of an electronic device suitable for implementing a network site selection prediction method according to an embodiment of the present disclosure is schematically shown.

[0133] like Fig.13 As shown, the electronic device 900 according to an embodiment of the present disclosure includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage part 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include an onboard memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0134] In RAM 903, various programs and data required for the operation of electronic device 900 are stored. Processor 901, ROM 902 and RAM 903 are connected to each other via bus 904. Processor 901 performs various operations of the method flow according to the embodiment of the present disclosure by executing the program in ROM 902 and / or RAM 903. It should be noted that the program can also be stored in one or more memories other than ROM 902 and RAM 903. Processor 901 can also perform various operations of the method flow according to the embodiment of the present disclosure by executing the program stored in the one or more memories.

[0135] According to an embodiment of the present disclosure, the electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to the bus 904. The electronic device 900 may further include one or more of the following components connected to the I / O interface 905: an input portion 906 including a keyboard, a mouse, etc.; an output portion 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage portion 908 including a hard disk, etc.; and a communication portion 909 including a network interface card such as a LAN card, a modem, etc. The communication portion 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the I / O interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed, so that a computer program read therefrom is installed into the storage portion 908 as needed.

[0136] The present disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or may exist independently without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present disclosure is implemented.

[0137] According to an embodiment of the present disclosure, a computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, an apparatus, or a device. For example, according to an embodiment of the present disclosure, a computer-readable storage medium may include the ROM 902 and / or RAM 903 described above and / or one or more memories other than ROM 902 and RAM 903.

[0138] The embodiment of the present disclosure also includes a computer program product, which includes a computer program, and the computer program contains program code for executing the method shown in the flowchart. When the computer program product is run in a computer system, the program code is used to enable the computer system to implement the method provided by the embodiment of the present disclosure.

[0139] The above functions defined in the system / device of the embodiment of the present disclosure are performed when the computer program is executed by the processor 901. According to the embodiment of the present disclosure, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0140] In one embodiment, the computer program may be based on a tangible storage medium such as an optical storage device, a magnetic storage device, etc. In another embodiment, the computer program may also be transmitted and distributed in the form of a signal on a network medium, and downloaded and installed through the communication part 909, and / or installed from a removable medium 911. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0141] In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 909, and / or installed from the removable medium 911. When the computer program is executed by the processor 901, the above functions defined in the system of the embodiment of the present disclosure are performed. According to the embodiment of the present disclosure, the system, device, means, module, unit, etc. described above can be implemented by a computer program module.

[0142] According to an embodiment of the present disclosure, the program code for executing the computer program provided by the embodiment of the present disclosure can be written in any combination of one or more programming languages. Specifically, these computing programs can be implemented using high-level process and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, Java, C++, python, "C" language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on the remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect through the Internet).

[0143] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0144] It will be appreciated by those skilled in the art that the features described in the various embodiments and / or claims of the present disclosure may be combined and / or combined in a variety of ways, even if such combinations and / or combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments and / or claims of the present disclosure may be combined and / or combined in a variety of ways without departing from the spirit and teachings of the present disclosure. All of these combinations and / or combinations fall within the scope of the present disclosure.

[0145] The embodiments of the present disclosure are described above. However, these embodiments are only for illustrative purposes and are not intended to limit the scope of the present disclosure. Although the embodiments are described above separately, this does not mean that the measures in the various embodiments cannot be used in combination to advantage. The scope of the present disclosure is defined by the attached claims and their equivalents. Without departing from the scope of the present disclosure, those skilled in the art may make a variety of substitutions and modifications, which should all fall within the scope of the present disclosure.

Claims

1. A network site selection prediction method, characterized in that: include: Acquire a target image containing a vehicle to be identified, wherein the target image includes m sub-images corresponding to the same geographical range at different time points, wherein m is an integer greater than or equal to 2; Performing vehicle recognition on the target image based on the pre-trained vehicle recognition model to obtain a vehicle recognition result, wherein the vehicle recognition result includes a vehicle position, a vehicle category, and a quantity of each type of vehicle corresponding to each sub-image; Acquiring vehicle flow data based on the vehicle recognition result; and Based on the traffic flow data, network site selection prediction is performed. The vehicle recognition model is obtained by training the YOLOv3 network improved by the deep residual module; The performing vehicle recognition on the target image based on the pre-trained vehicle recognition model to obtain a vehicle recognition result comprises: Processing the target image based on a feature extraction network to obtain a first feature image, wherein the first feature image includes k groups of first sub-feature images of different target sizes, where k is an integer greater than or equal to 6; wherein the feature extraction network includes a residual module and a convolutional layer, wherein the number of residual module groups is n, n satisfies k≤n and n is an integer, the number of convolutional layers is associated with the number of residual module groups, and the convolutional layers include convolutional layers that meet the target field of view; Inputting the first feature image into a feature fusion network to obtain a second feature image; Inputting the second feature image into a prediction layer network to obtain a predicted feature image; Generate a priori target frame based on the predicted feature image; and The priori target frame is processed based on a non-maximum suppression algorithm to obtain a vehicle recognition result.

2. The method according to claim 1, wherein: Inputting the first feature image into a feature fusion network to obtain a second feature image comprises: The kth group of first sub-feature maps obtained from the feature extraction network is marked as the first group of second sub-feature maps; After performing a 2-fold upsampling operation on the j-th group of second sub-feature maps, the maps are fused in a cascade manner with the kj-th group of first sub-feature maps output by the feature extraction network to obtain the s-th group of second sub-feature maps, where j traverses integer values ​​in [1, k-1] in ascending order, and s satisfies s=j+1 and is an integer; and A second feature image is obtained based on the 1st to kth groups of second sub-feature images.

3. The method according to claim 1, wherein: Processing the target image based on the feature extraction network to obtain the first feature image includes: Inputting the target image into the feature extraction network, and obtaining a first set of output images after being processed by the convolution layer; Processing the first group of output images based on the residual module to obtain a second group of output images; Alternatingly performing the convolutional layer processing step and the residual module processing step until the nth group of output images is obtained; and A first feature image is acquired based on the second to nth groups of output images.

4. The method according to claim 1, wherein: Processing the target image based on the feature extraction network to obtain the first feature image further includes: Discard output images that do not meet the target size.

5. The method according to claim 3, wherein: The acquiring of the first feature image based on the second to nth groups of output images further comprises: The second to the nkth groups of output images are discarded to obtain k groups of first sub-feature maps of different target sizes.

6. The method according to claim 1, wherein: Before performing vehicle recognition, the method further includes a step of preprocessing the target image, wherein the preprocessing includes adjusting the size of the target image so that the adjusted target image can output a first feature image of the target size after being processed by a feature extraction network.

7. The method according to claim 1, wherein: The first feature image includes 6 groups of first sub-feature images, and the sizes of the first sub-feature images are 64×64, 32×32, 16×16, 8×8, 4×4, and 2×2 respectively.

8. The method according to claim 6, wherein: The size of the target image is 512×512.

9. The method according to claim 1, wherein: The convolution layer does not include a 1×1 convolution layer.

10. The method according to claim 1, wherein: The geographic range includes q target areas, q is an integer greater than or equal to 2, and there is a mapping relationship between the vehicle position and the target area. The acquiring of the vehicle flow data based on the vehicle recognition result further includes: Count the number of each type of vehicles in the p-th target area in the sub-image at the i-th time point, marked as the first traffic flow data, where p satisfies 1≤p≤q and i satisfies 1≤i≤m; Acquire second traffic flow data based on the first traffic flow data of the sub-images at the m different time points, wherein the second traffic flow data includes the first traffic flow data of the sub-image at each time point; The second vehicle flow data is processed based on a preset algorithm to obtain the vehicle flow data, wherein a mapping relationship exists between the vehicle flow data and a target area.

11. The method according to claim 10, wherein: The vehicle flow data at least includes automobile flow data.

12. The method according to claim 10, wherein: The network site selection prediction based on the traffic flow data includes: Sort the traffic flow data corresponding to the q target areas in descending order; Obtain network site selection prediction results based on the sorting results.

13. A network site selection prediction device, comprising: An acquisition module is configured to acquire a target image containing a vehicle to be identified, wherein the target image includes m sub-images corresponding to the same geographical range at different time points, wherein m is an integer greater than or equal to 2; A recognition module is configured to perform vehicle recognition on the target image based on a pre-trained vehicle recognition model to obtain a vehicle recognition result, wherein the vehicle recognition result includes a vehicle position, a vehicle category, and a number of each type of vehicle corresponding to each sub-image, and the vehicle recognition model is obtained by training a YOLOv3 network improved by a deep residual module; A first processing module, configured to obtain vehicle flow data based on the vehicle recognition result; A second processing module is configured to perform network site selection prediction based on the vehicle flow data; The recognition module is further configured to process the target image based on a feature extraction network to obtain a first feature image, wherein the first feature image includes k groups of first sub-feature images of different target sizes, and k is an integer greater than or equal to 6; input the first feature image into a feature fusion network to obtain a second feature image; input the second feature image into a prediction layer network to obtain a predicted feature image; generate a priori target frame based on the predicted feature image; and process the prior target frame based on a non-maximum suppression algorithm to obtain a vehicle recognition result; wherein the feature extraction network includes a residual module and a convolutional layer, wherein the number of residual module groups is n, n satisfies k≤n and n is an integer, the number of convolutional layers is associated with the number of residual module groups, and the convolutional layers include convolutional layers that satisfy the target field of view.

14. An electronic device comprising: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors execute the method according to any one of claims 1 to 12.

15. A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, causes the processor to execute the method according to any one of claims 1 to 12.

16. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 12 is implemented.

Images