Method and system for target detection based on monitoring scenarios

Abstract

The application discloses a target detection method and system based on a monitoring scene. The method comprises the following steps: inputting target multi-external node data set into a feature extraction network to obtain target multi-external node feature map set; inputting each target multi-external node feature map into a target classification network for learning to obtain the predicted target category of each target; inputting each target multi-external node feature map into an external center point network for learning to obtain the predicted graph center point; inputting each target multi-external node feature map into a radial radiation distance network for learning to obtain the predicted radial radiation distance from each graph center point to each external node; inputting each target multi-external node feature map into a polar coordinate angle network for learning to obtain the predicted external node angle between adjacent external nodes; and performing model training and optimization according to the predicted target category, the predicted graph center point, the predicted radial radiation distance and the predicted external node angle.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a target detection method and system based on a monitoring scene. Background Art

[0002] In recent years, high-altitude video technology has developed rapidly. By installing video surveillance cameras on the roadside to collect data and using visual algorithms to collect data, various tasks such as vehicle detection, license plate recognition, vehicle tracking, pedestrian detection, and parking line detection can be achieved. This is conducive to more refined and accurate management of roadside parking. In addition, it can provide data for the deployment and development of vehicle-road collaborative technology, thereby having a positive promoting effect on various aspects of urban traffic management, driving safety, etc.

[0003] Currently, scene analysis in surveillance video primarily relies on two-dimensional object detection algorithms, which use rectangular boxes to describe the positions of multiple objects in the image. However, this method is relatively crude, describing only the approximate location and size of the object without fully capturing its shape, resulting in low object detection accuracy. Summary of the Invention

[0004] The purpose of the present invention is to solve the technical problem of low detection accuracy caused by the traditional method of using two-dimensional target detection. To achieve the above purpose, the present invention provides a target detection method and system based on monitoring scenes.

[0005] The present invention provides a target detection method based on a monitoring scenario, comprising:

[0006] Construct a target multi-external node dataset of surveillance scene images;

[0007] Inputting the target multi-external node data set into a feature extraction network for feature extraction to obtain a target multi-external node feature atlas;

[0008] Inputting each target multi-external node feature map in the target multi-external node feature map set into the target classification network for learning to obtain the predicted target category of each target;

[0009] Inputting each target multi-external node feature graph into the external center point network for learning, and obtaining the center point of the predicted graph formed by each multi-external node;

[0010] Inputting each target multi-external node feature graph into a radial radiation distance network for learning, and obtaining a predicted radial radiation distance from the center point of the graph to each external node;

[0011] Inputting each target multi-external node feature map into a polar coordinate angle network for learning to obtain predicted external node angles between adjacent external nodes;

[0012] According to the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumscribed node angle, a target multi-circumscribed detection model is trained and optimized to obtain a trained target multi-circumscribed detection model; wherein the feature extraction network, the target classification network, the circumscribed center point network, the radial radiation distance network, and the polar coordinate angle network form the target multi-circumscribed detection model;

[0013] The trained target multi-external detection model is used to perform target detection on the monitored scene image to obtain the category, position and shape of each target.

[0014] In one embodiment, after inputting each target multi-external node feature map into a polar coordinate angle network for learning and obtaining predicted external node angles between adjacent external nodes, the method further includes:

[0015] According to the Hungarian matching algorithm, the predicted target category, the predicted graphic center point, the predicted radial radiation distance, the predicted external node angle, the target category, the graphic center point, the radial radiation distance, and the external node angle are matched to obtain multiple matching pairs.

[0016] In one embodiment, constructing a target multi-external node dataset of a monitoring scene image includes:

[0017] Obtaining the target category, external node coordinates, graphic center point, radial radiation distance, and external node angle of each target in the monitoring scene image;

[0018] Get the expansion points on the line segment formed by adjacent external nodes;

[0019] Calculating the extended radial radiation distance between the center point of the graphic and the extended point;

[0020] Calculating an extended external node angle between the external node and the extended point;

[0021] The target category, the external node coordinates, the graphic center point, the radial radiation distance, the external node angle, the expansion point, the expanded radial radiation distance, and the expanded external node angle form the target multi-external node data set.

[0022] In one embodiment, the training and optimization of the target multi-circumference detection model based on the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumference node angle to obtain the trained target multi-circumference detection model includes:

[0023] A center point loss function is constructed based on the predicted center point of the graphic and the center point of the graphic. The center point loss function is:

[0024]

[0025] Where W represents the width of the graph formed by multiple external nodes, H represents the height of the graph formed by the multiple external nodes, l1 represents the L1 loss function, C x with C y Indicates the X-coordinate value and Y-coordinate value of the center point of the graphic, and Indicates the X-coordinate value and Y-coordinate value of the center point of the predicted graph.

[0026] In one embodiment, the training and optimizing of the target multi-circumference detection model based on the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumference node angle to obtain the trained target multi-circumference detection model further includes:

[0027] A radial radiation distance loss function is constructed according to the predicted radial radiation distance, the radial radiation distance, and the expanded radial radiation distance. The radial radiation distance loss function is:

[0028]

[0029] Among them, r i represents the radial radiation distance corresponding to the external node i or the extended radial radiation distance corresponding to the extended point i, Represents the predicted radial radiation distance corresponding to the external node i or the extension point i.

[0030] The present invention provides a target detection system based on a monitoring scenario, comprising:

[0031] A data construction module is used to construct a target multi-external node dataset for monitoring scene images;

[0032] A feature extraction module is used to input the target multi-external node data set into a feature extraction network to perform feature extraction and obtain a target multi-external node feature atlas;

[0033] A category learning module is used to input each target multi-external node feature map in the target multi-external node feature map set into a target classification network for learning to obtain a predicted target category for each target;

[0034] A center point learning module is used to input each target multi-external node feature graph into the external center point network for learning, and obtain the center point of the predicted graph formed by each multi-external node;

[0035] A distance learning module is used to input each target multi-external node feature graph into a radial radiation distance network for learning, and obtain a predicted radial radiation distance from the center point of the graph to each external node;

[0036] An angle learning module, used for inputting each target multi-external node feature map into a polar coordinate angle network for learning, and obtaining predicted external node angles between adjacent external nodes;

[0037] A model training module is used to train and optimize the target multi-circumference detection model based on the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumscribed node angle to obtain a trained target multi-circumference detection model; wherein the feature extraction network, the target classification network, the circumscribed center point network, the radial radiation distance network, and the polar coordinate angle network form the target multi-circumference detection model;

[0038] The target detection module is used to perform target detection on the monitored scene image to be tested based on the trained target multi-external detection model to obtain the category, position and shape of each target.

[0039] In one embodiment, the system further comprises:

[0040] A matching module is used to match the predicted target category, the predicted graphic center point, the predicted radial radiation distance, the predicted circumscribed node angle, the target category, the graphic center point, the radial radiation distance, and the circumscribed node angle according to the Hungarian matching algorithm to obtain multiple matching pairs.

[0041] In one embodiment, the data construction module includes:

[0042] A data acquisition module, configured to acquire the target category, external node coordinates, graphic center point, radial radiation distance, and external node angle of each target in the monitoring scene image;

[0043] An expansion data acquisition module, used to acquire expansion points on a line segment formed by adjacent external nodes;

[0044] a distance calculation module configured to calculate an extended radial radiation distance of the extended point from the graphic center point;

[0045] an angle calculation module configured to calculate an extended circumscribed node angle between the circumscribed node and the extended point;

[0046] a data forming module configured to form the target multi-circumscribed node dataset from the target category, the circumscribed node coordinates, the graphic center point, the radial radiation distance, the circumscribed node angle, the extended point, the extended radial radiation distance, and the extended circumscribed node angle.

[0047] In one embodiment, the model training module comprises:

[0048] a center point loss function construction module configured to construct a center point loss function from the predicted graphic center point and the graphic center point, the center point loss function being:

[0049]

[0050] wherein W represents a width of the multi-circumscribed node formed graphic, H represents a height of the multi-circumscribed node formed graphic, l1 represents an L1 loss function, C x and C y represent X coordinate and Y coordinate values of the graphic center point, and represent X coordinate and Y coordinate values of the predicted graphic center point.

[0051] In one embodiment, the model training module further comprises:

[0052] a distance loss function construction module configured to construct a radial radiation distance loss function from the predicted radial radiation distance, the radial radiation distance, and the extended radial radiation distance, the radial radiation distance loss function being:

[0053]

[0054] wherein r i represents the radial radiation distance corresponding to the circumscribed node i or the extended radial radiation distance corresponding to the extended point i, represents the predicted radial radiation distance corresponding to the circumscribed node i or the extended point i.

[0055] In the above-mentioned target detection method and system based on the monitoring scene, a polygon detection method is used to detect the target, and the outer contour of the target is described by using multiple external nodes. Compared with the two-dimensional rectangular frame in the traditional method, the present invention detects the position and shape of the target from multiple dimensions such as the center point of the graphic, the radial radiation distance, and the external node angle, which can more accurately describe the position and shape of the target and improve the accuracy of the target detection result. At the same time, compared with the instance segmentation algorithm in the traditional method, less computing power is required, and real-time calculations can be performed on the application device side, which saves the computing power calculation process and improves the efficiency of target detection. By outputting the target through the target detection method provided by the present invention, the high-position video scene on the roadside can be analyzed more accurately, which is more helpful for the management of roadside parking. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic flow chart of the steps of the target detection method based on the monitoring scene provided by the present invention.

[0057] Figure 2 It is a structural diagram of the target detection system based on monitoring scenarios provided by the present invention. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

[0059] See Figure 1 The present invention provides a target detection method based on a monitoring scenario, comprising:

[0060] S10, constructing a target multi-external node dataset of the monitoring scene image;

[0061] S20, inputting the target multi-external node dataset into a feature extraction network to perform feature extraction, and obtaining a target multi-external node feature atlas;

[0062] S30, inputting each target multi-external node feature map in the target multi-external node feature map set into the target classification network for learning, to obtain a predicted target category for each target;

[0063] S40, inputting each target multi-external node feature graph into the external center point network for learning, and obtaining the center point of the predicted graph formed by each multi-external node;

[0064] S50, inputting each target multi-external node feature graph into a radial radiation distance network for learning, and obtaining a predicted radial radiation distance from the center point of the graph to each external node;

[0065] S60, inputting each target multi-external node feature map into a polar coordinate angle network for learning to obtain predicted external node angles between adjacent external nodes;

[0066] S70, training and optimizing a target multi-circumference detection model based on the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumference node angle to obtain a trained target multi-circumference detection model; wherein the feature extraction network, the target classification network, the circumference center point network, the radial radiation distance network, and the polar coordinate angle network form the target multi-circumference detection model;

[0067] S80: Perform target detection on the monitored scene image to be tested based on the trained target multi-external detection model to obtain the category, position and shape of each target.

[0068] In this embodiment, the monitoring scene image is captured by using a video surveillance camera installed on the roadside. The monitoring scene image covers data under different monitoring camera installation scenarios, different monitoring angles, and different weather conditions. The targets in the monitoring scene image, such as vehicles and pedestrians, are annotated using a polygon annotation tool. The polygon nodes and annotation categories of each target instance are recorded. For polygon nodes, the recording is carried out in a clockwise direction, O = {poly = [(x1, y1), (x2, y2), (x3, y3), ..., (x k ,y k )], cls}, forming a target multi-external node dataset. poly represents the coordinates of the multiple nodes that constitute the target polygon. The number of polygon nodes K is not fixed; different targets can use different numbers of nodes to describe the target polygon. cls represents the category, which can include pedestrians, vehicles, non-motor vehicles, and other target categories.

[0069] The feature extraction network can be a ResNet-18 backbone network to extract features from image frames, and the input image size is the resolution size of the original video. Feature extraction is performed through the feature extraction network for subsequent learning of the polygonal target detection network. The function of the target classification network is to predict and output the category of the target. The target classification network can adopt a conventional classification network with multiple convolutional layers. The target classification network regresses and outputs multiple external nodes of the target. The external center point network adopts a full convolutional neural network model to learn the center point. The external center point can be understood as the center of mass of the polygon formed by the external nodes. The radial radiation distance network adopts a full convolutional neural network model to learn the distance. The radial radiation distance is the distance from the center point of each graphic to each external node, expressed as:

[0070]

[0071] Among them, ri denotes the radial radiation distance from the center point of the figure to the i-th circumscribed node, C x , C y denote the x and y coordinates of the center point of the figure, k ix , k iy denote the x and y coordinates of the i-th circumscribed node. The polar coordinate angle network adopts a full convolutional neural network model to learn the angle value. The polar coordinate angle represents the angle value between adjacent circumscribed nodes, for example, the center point of the polygon is set as the center of the circle, the circle is divided into 360°, the x coordinate zero point is 0°, and the 360° is divided counterclockwise.

[0072] The target loss function is constructed by the prediction target category, the prediction figure center point, the prediction radial radiation distance and the prediction circumscribed node angle predicted by each network structure. The feature extraction network, the target classification network, the circumscribed center point network, the radial radiation distance network and the polar coordinate angle network are trained to obtain the trained feature extraction network, the trained target classification network, the trained circumscribed center point network, the trained radial radiation distance network and the trained polar coordinate angle network. The category of each target can be detected by the trained target classification network. The figure center point of each target can be detected by the trained circumscribed center point network. The distance from the figure center point of each target to each circumscribed node can be detected by the trained radial radiation distance network, and then the coordinate position of each circumscribed node is obtained. The angle between two adjacent circumscribed nodes of each target can be detected by the trained polar coordinate angle network. The figure center point, the radial radiation distance and the circumscribed node angle of each target can indicate the position and shape of each target.

[0073] The target detection method based on the monitoring scene provided by the application adopts a polygon detection method to detect the target, describes the outer contour of the target by using multiple circumscribed nodes, detects the position and shape of the target from multiple dimensions of the figure center point, the radial radiation distance and the circumscribed node angle relative to the two-dimensional rectangular frame in the traditional method, can more accurately describe the position and shape of the target, and improves the accuracy of the target detection result. At the same time, compared with the instance segmentation algorithm in the traditional method, less computing power is required, real-time operation can be performed on the application device end, the process of computing power calculation is saved, and the efficiency of target detection is improved. The output of the target detection method provided by the application can more accurately analyze the roadside high-position video scene, and is more helpful for the management of roadside parking.

[0074] In one embodiment, at S60, after inputting each target multi-external node feature map into a polar coordinate angle network for learning and obtaining predicted external node angles between adjacent external nodes, the method further includes:

[0075] S610 , matching the predicted target category, the predicted graphic center point, the predicted radial radiation distance, the predicted circumscribed node angle, the target category, the graphic center point, the radial radiation distance, and the circumscribed node angle according to the Hungarian matching algorithm to obtain a plurality of matching pairs.

[0076] In this embodiment, the Hungarian matching algorithm is a graph theory algorithm for finding the maximum match. It matches multiple predicted objects of the target in the image with multiple ground-truth targets in the annotated data to form multiple matching pairs. This can also be understood as matching predicted target categories with target categories, matching predicted graphic centers with graphic centers, matching predicted radial distances with radial distances, and matching predicted circumscribed node angles with circumscribed node angles to form multiple matching pairs. These matching pairs are used in the subsequent steps to construct the target loss function to achieve model training and optimization.

[0077] In one embodiment, S10, constructing a target multi-external node dataset of a monitoring scene image includes:

[0078] S110, obtaining the target category, external node coordinates, graphic center point, radial radiation distance, and external node angle of each target in the monitoring scene image;

[0079] S120, obtaining expansion points on a line segment formed by adjacent external nodes;

[0080] S130, calculating the extended radial radiation distance between the center point of the graphic and the extended point;

[0081] S140, calculating the extended external node angle between the external node and the extended point;

[0082] S150 , forming a target multi-external node data set by the target category, external node coordinates, graphic center point, radial radiation distance, external node angle, expansion point, expanded radial radiation distance, and expanded external node angle.

[0083] In this embodiment, multiple external nodes form a polygon. The polygon is divided into multiple triangles, and the area of ​​each triangle is calculated to obtain the total area of ​​the polygon. The centroid coordinates of each triangle constituting the polygon are calculated and multiplied by the area of ​​the corresponding triangle to obtain the sum of the products of the centroids and areas of all triangles. The centroid of the polygon is calculated based on the sum and the area of ​​the polygon, which is expressed as:

[0084]

[0085] Among them, C x , C y represent the x and y coordinates of the polygon centroid, C ix represents the x-coordinate of the centroid of each triangle, A i Represents the area of ​​a triangle, i represents the polygon is divided into i triangles, used to calculate the area of ​​the polygon. The x coordinate of the triangle centroid is calculated as: x1, x2, and x3 represent the x-coordinates of the three vertices of the triangle.

[0086] C iy represents the y coordinate of the triangle's centroid, y1, y2, and y3 represent the y coordinates of the three vertices of the triangle. The centroid of the polygon is the center of the shape.

[0087] For a polygonal object, a total of K external nodes are used for annotation. The number of these K external nodes is expanded to form multiple expansion points. Adjacent external nodes A and B are two adjacent polygon nodes, O is the center of mass of the polygon, and an expansion point C is added on the line segment formed between A and B.

[0088] Based on the principle of similar triangles, the coordinates of point C, the radial distance of OC, the angle formed by OA and OC, and the angle formed by OC and OB are calculated. The coordinates of point C are the coordinates of the expansion point. The radial distance of OC is the expanded radial distance. The angles formed by OA and OC, and the angles formed by OC and OB, are the expanded circumscribed node angles. The annotation data includes the target category, circumscribed node coordinates, the center point of the shape, the radial distance, the circumscribed node angle, the expansion point, the expanded radial distance, and the expanded circumscribed node angle.

[0089] In this embodiment, the target multi-external node dataset is enriched by adding multiple expansion points. When constructing the target loss function for the labeled data and the model prediction results, the same number of nodes can be used, solving the problem of different numbers of labeled nodes corresponding to different targets, and the loss function can be directly calculated for the model.

[0090] In one embodiment, S70, training and optimizing the target multi-circumference detection model based on the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumference node angle to obtain a trained target multi-circumference detection model includes:

[0091] S710: Construct a center point loss function based on the predicted center point of the graphic and the center point of the graphic. The center point loss function is:

[0092]

[0093] Among them, W represents the width of the graph formed by multiple external nodes, H represents the height of the graph formed by multiple external nodes, l1 represents the L1 loss function, C x with C y Indicates the X and Y coordinate values ​​of the center point of the graphic. and Indicates the X and Y coordinate values ​​of the center point of the predicted graph.

[0094] In one embodiment, S70, training and optimizing the target multi-circumference detection model based on the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumference node angle to obtain a trained target multi-circumference detection model further includes:

[0095] S720: Construct a radial radiation distance loss function based on the predicted radial radiation distance, the radial radiation distance, and the expanded radial radiation distance. The radial radiation distance loss function is:

[0096]

[0097] Among them, r i Indicates the radial radiation distance corresponding to the external node i or the extended radial radiation distance corresponding to the extended point i, Indicates the predicted radial radiation distance corresponding to the external node i or the expansion point i.

[0098] See Figure 2 The present invention provides a target detection system 100 based on a monitoring scene. The target detection system 100 based on a monitoring scene includes a data construction module 10, a feature extraction module 20, a category learning module 30, a center point learning module 40, a distance learning module 50, an angle learning module 60, a model training module 70 and a target detection module 80. The data construction module 10 is used to construct a target multi-external node dataset of the monitoring scene image. The feature extraction module 20 is used to input the target multi-external node dataset into a feature extraction network for feature extraction to obtain a target multi-external node feature graph set. The category learning module 30 is used to input each target multi-external node feature graph in the target multi-external node feature graph set into a target classification network for learning to obtain a predicted target category for each target.

[0099] The center point learning module 40 is used to input each target multi-external node feature map into the external center point network for learning, thereby obtaining the predicted center point of each multi-external node. The distance learning module 50 is used to input each target multi-external node feature map into the radial radiation distance network for learning, thereby obtaining the predicted radial radiation distance from the center point of the graph to each external node. The angle learning module 60 is used to input each target multi-external node feature map into the polar coordinate angle network for learning, thereby obtaining the predicted external node angles between adjacent external nodes.

[0100] The model training module 70 is used to train and optimize the target multi-external detection model based on the predicted target category, predicted graphic center point, predicted radial radiation distance, and predicted circumscribed node angle, thereby obtaining a trained target multi-external detection model. The target multi-external detection model is composed of a feature extraction network, a target classification network, a circumscribed center point network, a radial radiation distance network, and a polar coordinate angle network. The target detection module 80 is used to perform target detection on the monitored scene image using the trained target multi-external detection model, obtaining the category, location, and shape of each target.

[0101] In this embodiment, the relevant description of the data construction module 10 may refer to the relevant description of S10 in the above embodiment. The relevant description of the feature extraction module 20 may refer to the relevant description of S20 in the above embodiment. The relevant description of the category learning module 30 may refer to the relevant description of S30 in the above embodiment. The relevant description of the center point learning module 40 may refer to the relevant description of S40 in the above embodiment. The relevant description of the distance learning module 50 may refer to the relevant description of S50 in the above embodiment. The relevant description of the angle learning module 60 may refer to the relevant description of S60 in the above embodiment. The relevant description of the model training module 70 may refer to the relevant description of S70 in the above embodiment. The relevant description of the target detection module 80 may refer to the relevant description of S80 in the above embodiment.

[0102] In one embodiment, the monitoring scene-based target detection system 100 further includes a matching module. The matching module is configured to match the predicted target category, the predicted graph center point, the predicted radial radiation distance, the predicted circumscribed node angle, the target category, the graph center point, the radial radiation distance, and the circumscribed node angle according to the Hungarian matching algorithm to obtain a plurality of matching pairs.

[0103] In this embodiment, the relevant description of the matching module may refer to the relevant description of S610 in the above embodiment.

[0104] In one embodiment, the data construction module 10 includes a data acquisition module, an expanded data acquisition module, a distance calculation module, an angle calculation module, and a data formation module. The data acquisition module is used to obtain the target category, circumscribed node coordinates, graph center point, radial radiation distance, and circumscribed node angle of each target in the monitored scene image. The expanded data acquisition module is used to obtain expanded points on the line segment formed by adjacent circumscribed nodes. The distance calculation module is used to calculate the expanded radial radiation distance between the graph center point and the expanded points.

[0105] The angle calculation module is used to calculate the extended external node angle between the external node and the expansion point. The data formation module is used to form a target multi-external node data set based on the target category, external node coordinates, graphic center point, radial radiation distance, external node angle, expansion point, extended radial radiation distance, and extended external node angle.

[0106] In this embodiment, the description of the data acquisition module can refer to the description of S110 in the above embodiment. The description of the expanded data acquisition module can refer to the description of S120 in the above embodiment. The description of the distance calculation module can refer to the description of S130 in the above embodiment. The description of the angle calculation module can refer to the description of S140 in the above embodiment. The description of the data formation module can refer to the description of S150 in the above embodiment.

[0107] In one embodiment, the model training module 70 includes a center point loss function construction module. The center point loss function construction module is used to construct a center point loss function based on the predicted center point of the graphic and the center point of the graphic. The center point loss function is:

[0108]

[0109] Among them, W represents the width of the graph formed by multiple external nodes, H represents the height of the graph formed by multiple external nodes, l1 represents the L1 loss function, C x with C y Indicates the X and Y coordinate values ​​of the center point of the graphic. and Indicates the X and Y coordinate values ​​of the center point of the predicted graph.

[0110] In this embodiment, the relevant description of the center point loss function construction module can refer to the relevant description of S710 in the above embodiment.

[0111] In one embodiment, the model training module 70 further includes a distance loss function construction module. The distance loss function construction module is used to construct a radial radiation distance loss function based on the predicted radial radiation distance, the radial radiation distance, and the expanded radial radiation distance. The radial radiation distance loss function is:

[0112]

[0113] wherein, r i denotes the predicted radial radiation distance corresponding to the circumscribed node i or the augmented node i. denotes the predicted radial radiation distance corresponding to the circumscribed node i or the augmented node i.

[0114] In this embodiment, the related description of the distance loss function construction module can refer to the related description of S720 in the above-mentioned embodiments.

[0115] In each of the above-mentioned embodiments, the specific order or hierarchy of the steps in the disclosed processes is an example. Based on design preference, it should be understood that the specific order or hierarchy of the steps in the processes can be rearranged without departing from the scope of protection of the present disclosure. The appended method claims list the elements of the various steps in an exemplary order, and are not intended to be limited to a particular order or hierarchy.

[0116] Those skilled in the art can also understand that the various illustrative logical blocks (illustrative logical blocks), modules and steps listed in the embodiments of the present application can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly show the interchangeability of hardware and software, the above-mentioned various illustrative components (illustrative components), modules and steps have been generally described their functions. Whether the functions are implemented by hardware or software depends on the specific application and design requirements of the overall system. Those skilled in the art can implement the functions in various ways for each specific application, but such implementation should not be understood as beyond the scope of protection of the embodiments of the present application.

[0117] The various illustrative logical blocks or modules described in the embodiments of the present application can be implemented or operated by a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field programmable gate array or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination of the above. The general-purpose processor can be a microprocessor, and optionally, the general-purpose processor can also be any conventional processor, controller, microcontroller or state machine. The processor can also be implemented by a combination of computing devices, such as a digital signal processor and a microprocessor, multiple microprocessors, one or more microprocessors in combination with a digital signal processor core, or any other similar configuration.

[0118] The steps of the methods or algorithms described in the embodiments of the present invention may be directly embedded in hardware, a software module executed by a processor, or a combination of the two. The software module may be stored in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. For example, the storage medium may be connected to the processor so that the processor can read information from the storage medium and write information to the storage medium. Alternatively, the storage medium may also be integrated into the processor. The processor and storage medium may be provided in an ASIC, which may be provided in a user terminal. Alternatively, the processor and storage medium may also be provided in different components in the user terminal.

[0119] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A target detection method based on a monitoring scene, characterized in that: include: Construct a target multi-external node dataset of surveillance scene images; Inputting the target multi-external node data set into a feature extraction network for feature extraction to obtain a target multi-external node feature atlas; Inputting each target multi-external node feature map in the target multi-external node feature map set into the target classification network for learning to obtain the predicted target category of each target; Inputting each target multi-external node feature graph into the external center point network for learning, and obtaining the center point of the predicted graph formed by each multi-external node; Inputting each target multi-external node feature graph into a radial radiation distance network for learning, and obtaining a predicted radial radiation distance from the center point of the graph to each external node; Inputting each target multi-external node feature map into a polar coordinate angle network for learning to obtain predicted external node angles between adjacent external nodes; According to the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumscribed node angle, a target multi-circumscribed detection model is trained and optimized to obtain a trained target multi-circumscribed detection model; wherein the feature extraction network, the target classification network, the circumscribed center point network, the radial radiation distance network, and the polar coordinate angle network form the target multi-circumscribed detection model; The trained target multi-external detection model is used to perform target detection on the monitored scene image to obtain the category, position and shape of each target.

2. The target detection method based on monitoring scene according to claim 1, characterized in that: After inputting each target multi-external node feature map into a polar coordinate angle network for learning and obtaining predicted external node angles between adjacent external nodes, the method further includes: The predicted target category, the predicted graphic center point, the predicted radial radiation distance, the predicted circumscribed node angle, the target category, the graphic center point, the radial radiation distance, and the circumscribed node angle are matched according to the Hungarian matching algorithm to obtain multiple matching pairs.

3. The target detection method based on monitoring scene according to claim 1, characterized in that: The target multi-external node dataset for constructing the monitoring scene image includes: Obtaining the target category, external node coordinates, graphic center point, radial radiation distance, and external node angle of each target in the monitoring scene image; Get the expansion points on the line segment formed by adjacent external nodes; Calculating the extended radial radiation distance between the center point of the graphic and the extended point; Calculating an extended external node angle between the external node and the extended point; The target category, the external node coordinates, the graphic center point, the radial radiation distance, the external node angle, the expansion point, the expanded radial radiation distance, and the expanded external node angle form the target multi-external node data set.

4. The target detection method based on monitoring scene according to claim 3 is characterized in that: The training and optimization of the target multi-circumference detection model according to the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumference node angle to obtain the trained target multi-circumference detection model includes: A center point loss function is constructed based on the predicted center point of the graphic and the center point of the graphic. The center point loss function is: Where W represents the width of the graph formed by multiple external nodes, H represents the height of the graph formed by the multiple external nodes, l1 represents the L1 loss function, C x with C y Indicates the X-coordinate value and Y-coordinate value of the center point of the graphic, and Indicates the X-coordinate value and Y-coordinate value of the center point of the predicted graph.

5. The target detection method based on monitoring scene according to claim 4, characterized in that: The step of training and optimizing the target multi-circumference detection model according to the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumference node angle to obtain a trained target multi-circumference detection model further includes: A radial radiation distance loss function is constructed according to the predicted radial radiation distance, the radial radiation distance, and the expanded radial radiation distance. The radial radiation distance loss function is: Among them, r i represents the radial radiation distance corresponding to the external node i or the extended radial radiation distance corresponding to the extended point i, represents the predicted radial radiation distance corresponding to the external node i or the expansion point i, and k represents the number of external nodes or the number of expansion points.

6. A target detection system based on a monitoring scene, characterized in that: include: A data construction module is used to construct a target multi-external node dataset for monitoring scene images; A feature extraction module is used to input the target multi-external node data set into a feature extraction network to perform feature extraction and obtain a target multi-external node feature atlas; A category learning module is used to input each target multi-external node feature map in the target multi-external node feature map set into a target classification network for learning to obtain a predicted target category for each target; A center point learning module is used to input each target multi-external node feature graph into the external center point network for learning, and obtain the center point of the predicted graph formed by each multi-external node; A distance learning module is used to input each target multi-external node feature graph into a radial radiation distance network for learning, and obtain a predicted radial radiation distance from the center point of the graph to each external node; An angle learning module, used for inputting each target multi-external node feature map into a polar coordinate angle network for learning, and obtaining predicted external node angles between adjacent external nodes; A model training module is used to train and optimize the target multi-circumference detection model based on the predicted target category, the predicted graphic center point, the predicted radial radiation distance, and the predicted circumscribed node angle to obtain a trained target multi-circumference detection model; wherein the feature extraction network, the target classification network, the circumscribed center point network, the radial radiation distance network, and the polar coordinate angle network form the target multi-circumference detection model; The target detection module is used to perform target detection on the monitored scene image to be tested based on the trained target multi-external detection model to obtain the category, position and shape of each target.

7. The target detection system based on monitoring scene according to claim 6, characterized in that: The system further comprises: A matching module is used to match the predicted target category, the predicted graphic center point, the predicted radial radiation distance, the predicted circumscribed node angle, the target category, the graphic center point, the radial radiation distance, and the circumscribed node angle according to the Hungarian matching algorithm to obtain multiple matching pairs.

8. The target detection system based on monitoring scene according to claim 6, characterized in that: The data construction module includes: A data acquisition module, configured to acquire the target category, external node coordinates, graphic center point, radial radiation distance, and external node angle of each target in the monitoring scene image; An expansion data acquisition module, used to acquire expansion points on a line segment formed by adjacent external nodes; A distance calculation module, configured to calculate an expanded radial radiation distance between the center point of the graphic and the expanded point; An angle calculation module, used for calculating the extended external node angle between the external node and the extended point; The data forming module is used to form the target multi-external node data set using the target category, the external node coordinates, the graphic center point, the radial radiation distance, the external node angle, the expansion point, the expanded radial radiation distance, and the expanded external node angle.

9. The target detection system based on monitoring scene according to claim 8, characterized in that: The model training module includes: A center point loss function construction module is used to construct a center point loss function based on the predicted graphic center point and the graphic center point. The center point loss function is: Where W represents the width of the graph formed by multiple external nodes, H represents the height of the graph formed by the multiple external nodes, l1 represents the L1 loss function, C x with C y Indicates the X-coordinate value and Y-coordinate value of the center point of the graphic, and Indicates the X-coordinate value and Y-coordinate value of the center point of the predicted graph.

10. The target detection system based on monitoring scene according to claim 9, characterized in that: The model training module also includes: A distance loss function construction module is used to construct a radial radiation distance loss function according to the predicted radial radiation distance, the radial radiation distance, and the expanded radial radiation distance. The radial radiation distance loss function is: Among them, r i represents the radial radiation distance corresponding to the external node i or the extended radial radiation distance corresponding to the extended point i, represents the predicted radial radiation distance corresponding to the external node i or the expansion point i, and k represents the number of external nodes or the number of expansion points.

Images