A smart traffic vehicle violation detection method and system

Abstract

This invention discloses a method and system for detecting vehicle violations in intelligent transportation, relating to the field of intelligent transportation. The method includes: performing receptive field fusion recognition using interactive historical monitoring images; stitching and fusing the background image of the target detection area based on multiple stitching and fusion routes and multiple receptive field fusion distribution fields; calling K monitoring devices within a preset detection window to collect images and environmental features of the target detection area; matching a first set of detection environmental features with semantic information from multiple detection scenarios, and mapping multiple stitching and fusion routes to obtain a first stitching and fusion route; fusing the K first detection images; and using a lightweight detection network layer to detect and recognize vehicle violations using the first detected fused images and the first region fused background. This method solves the technical problem of insufficient detection accuracy in existing vehicle violation detection systems, achieving a significant improvement in vehicle violation detection accuracy.

Description

Technical Field

[0001] The present application relates to the field of intelligent transportation, and in particular to a method and system for detecting vehicle violations in intelligent transportation. Background Art

[0002] In the wave of smart city construction, smart transportation system, as a key component, is of great significance in improving road safety, alleviating traffic congestion, optimizing traffic management, etc. Among them, traffic vehicle violation detection is an important part of the smart transportation system, which aims to achieve real-time monitoring and recording of traffic violations through automated means to assist traffic management departments in efficient law enforcement. At present, traffic vehicle violation detection mainly relies on fixed or mobile monitoring devices (such as cameras) deployed along the road. These monitoring devices capture video images and use image processing technology and machine learning algorithms to analyze vehicle behaviors in the images to identify violations (such as running red lights, driving on the line, illegal parking, etc.). However, due to the complexity of traffic scenes, such as weather changes, different lighting conditions, high vehicle density, severe occlusion and other factors, traditional violation detection methods are difficult to accurately and comprehensively detect violations, and the misjudgment and missed judgment rates are high.

[0003] Among the current related technologies, vehicle violation detection has the technical problem of insufficient detection accuracy. Summary of the invention

[0004] The present application provides a method and system for intelligent traffic vehicle violation detection, which uses receptive field fusion recognition to determine multiple receptive field fusion distribution fields, performs background image splicing and fusion and image collection of the target detection area, and uses a lightweight detection network layer to perform vehicle violation detection and recognition on the fused image, thereby achieving the technical effect of improving the accuracy of vehicle violation detection.

[0005] The present application provides a method for detecting traffic violations of smart traffic vehicles, which is applied to a smart traffic vehicle violation detection system, wherein the smart traffic vehicle violation detection system is communicatively connected to a monitoring device array, and the method comprises:

[0006] The historical monitoring images of the interactive monitoring device array are subjected to receptive field fusion identification to determine multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple detection scene semantic information and multiple splicing fusion routes, and the monitoring device array includes K monitoring devices, and the K monitoring devices include K device positioning identifiers; the background images of the target detection area are spliced ​​and fused according to the multiple splicing fusion routes and the multiple receptive field fusion distribution fields to obtain multiple regional fusion backgrounds; the K monitoring devices are called within a preset detection window to perform image acquisition and environmental feature acquisition on the target detection area to obtain K first detection images and a first detection environment feature set; the first detection environment feature set is matched with the multiple detection scene semantic information, and the first splicing fusion route is mapped according to the matching result and combined with the multiple splicing fusion routes to obtain a first splicing fusion route; the K first detection images are fused using the first splicing fusion route to obtain a first detection fusion image; the first detection fusion image and the first regional fusion background corresponding to the first splicing fusion route are subjected to vehicle violation detection and identification using a lightweight detection network layer to obtain a vehicle violation detection result.

[0007] In a possible implementation, the historical monitoring images of the interactive monitoring device array are subjected to receptive field fusion recognition, multiple receptive field fusion distribution fields are determined, and the following processing is performed:

[0008] The K monitoring devices of the interactive monitoring device array perform simultaneous sequential image collection at multiple historical time nodes to obtain multiple historical monitoring image sets, wherein each historical monitoring image set includes K historical monitoring images; collect multiple historical detection scene semantic information of the multiple historical monitoring image sets, perform semantic twin recognition on the multiple historical detection scene semantic information, obtain multiple aggregated detection scene semantic information sets and multiple historical monitoring image clusters; perform receptive field fusion recognition on the multiple historical monitoring image clusters in combination with the K device positioning identifiers to obtain multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple splicing fusion routes; perform feature fusion recognition on the multiple aggregated detection scene semantic information sets to obtain the multiple detection scene semantic information.

[0009] In a possible implementation, multiple historical detection scene semantic information of the multiple historical monitoring screen sets are collected, semantic twin recognition is performed on the multiple historical detection scene semantic information, multiple aggregated detection scene semantic information sets and multiple historical monitoring screen clusters are obtained, and the following processing is performed:

[0010] Construct a semantic twin recognition branch, wherein the semantic twin recognition branch is embedded with a twin network algorithm; randomly extract M historical detection scene semantic information from the multiple historical detection scene semantic information, and perform node authentication on the M historical detection scene semantic information. If the authentication is passed, the M historical detection scene semantic information is used as M identification nodes, wherein M is a positive integer greater than or equal to 3; respectively use the semantic twin recognition branch to perform twin recognition on the N historical detection scene semantic information and the M identification nodes, and divide the N historical detection scene semantic information into the identification node with the largest twin degree to obtain M aggregated detection scene semantic information, wherein the sum of M and N is the total number of multiple historical detection scene semantic information; use the M aggregated detection scene semantic information as multiple aggregated detection scene semantic information, and map to obtain multiple historical monitoring screen clusters.

[0011] In a possible implementation, M pieces of historical detection scene semantic information are randomly extracted from the plurality of historical detection scene semantic information, and node authentication is performed on the M pieces of historical detection scene semantic information, and the following processing is performed:

[0012] The M historical detection scene semantic information are enumerated and combined in pairs to obtain multiple historical detection scene semantic information combinations; the semantic twin recognition branch is used to perform twin degree recognition on the multiple historical detection scene semantic information combinations to obtain multiple combined twin degrees; when the multiple combined twin degrees are all less than or equal to the preset twin degree threshold, the authentication is passed.

[0013] In a possible implementation, the receptive field fusion recognition is performed on the multiple historical monitoring screen clusters in combination with the K device location identifiers to obtain multiple receptive field fusion distribution fields, and the following processing is performed:

[0014] A first historical monitoring screen cluster is randomly extracted from the multiple historical monitoring screen clusters, and then a first historical monitoring screen set is randomly extracted from the first historical monitoring screen cluster; the overlapping area of ​​the first historical monitoring screen set is identified in combination with the K device positioning identifiers to obtain multiple identified historical screen overlapping areas; a first screen intersection set of the first identified historical screen overlapping area of ​​the multiple identified historical screen overlapping areas is extracted, and any one of the screen intersections in the first screen intersection set is used as a starting point to perform transition gradient identification in the first identified historical screen overlapping area according to a preset iteration step to obtain a first stage splicing and fusion point; then the first stage splicing and fusion point is used as a stage starting point to perform multiple transition gradient identifications to obtain a first splicing and fusion sub-route of the first identified historical screen overlapping area; receptive field fusion identification is performed based on the multiple historical monitoring screen clusters to obtain multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple splicing and fusion routes.

[0015] In a possible implementation, the following processing is performed:

[0016] In combination with the K device positioning identifiers, receptive field fusion identification is performed on the overlapping areas of the multiple identified historical images to obtain a first spliced ​​fusion sub-route set of the first historical monitoring image set; in combination with the K device positioning identifiers, receptive field fusion identification is performed on the first historical monitoring image cluster to obtain multiple spliced ​​fusion sub-route sets; the multiple spliced ​​fusion sub-route sets are fitted with the overlapping area to obtain a first spliced ​​fusion route of the first historical monitoring image cluster; the target detection area is divided by using the first spliced ​​fusion route in combination with the K monitoring ranges of the K monitoring devices to obtain a first receptive field fusion distribution field, wherein the first receptive field fusion distribution field corresponds to a detection scene semantic information; in combination with the K device positioning identifiers, receptive field fusion identification is performed on the multiple historical monitoring image clusters to obtain multiple receptive field fusion distribution fields.

[0017] In a possible implementation, the following processing is performed:

[0018] After multiple transition gradient identifications, when the screen intersection farthest from the starting point in the first screen intersection set is reached, identification is stopped to obtain multiple stage splicing and fusion points; the first stage splicing and fusion points and the multiple stage splicing and fusion points are sequentially connected in the identification order to obtain the first splicing and fusion sub-route of the overlapping area of ​​the first identified historical screen.

[0019] The present application also provides a smart traffic vehicle violation detection system, comprising:

[0020] A module for determining multiple receptive field fusion distribution fields, the module for determining multiple receptive field fusion distribution fields is used to perform receptive field fusion identification on historical monitoring images of an interactive monitoring device array, and determine multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple detection scene semantic information and multiple splicing fusion routes, the monitoring device array includes K monitoring devices, and the K monitoring devices include K device positioning identifiers; a background image splicing fusion module, the background image splicing fusion module is used to perform background image splicing fusion of a target detection area according to the multiple splicing fusion routes and the multiple receptive field fusion distribution fields, and obtain multiple regional fusion backgrounds; a target detection area information acquisition module, the target detection area information acquisition module is used to call the K monitoring devices within a preset detection window to perform image acquisition on the target detection area and environmental feature collection to obtain K first detection images and a first detection environment feature set; a first splicing and fusion route acquisition module, the first splicing and fusion route acquisition module is used to match the first detection environment feature set with the multiple detection scene semantic information, and map the first splicing and fusion route according to the matching result to obtain a first splicing and fusion route; a first detection fusion image acquisition module, the first detection fusion image acquisition module is used to fuse the K first detection images using the first splicing and fusion route to obtain a first detection fusion image; a vehicle violation detection and recognition module, the vehicle violation detection and recognition module is used to use a lightweight detection network layer to perform vehicle violation detection and recognition on the first detection fusion image and the first regional fusion background corresponding to the first splicing and fusion route to obtain a vehicle violation detection result.

[0021] The present application proposes a method and system for detecting traffic violations of intelligent traffic vehicles. First, the historical monitoring images of the interactive monitoring device array are used for receptive field fusion recognition to determine multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple detection scene semantic information and multiple splicing fusion routes, the monitoring device array includes K monitoring devices, and the K monitoring devices include K device positioning identifiers. Then, the background images of the target detection area are spliced ​​and fused according to the multiple splicing fusion routes and the multiple receptive field fusion distribution fields to obtain multiple regional fusion backgrounds, and then the K monitoring devices are called within the preset detection window to draw the target detection area. The first detection image and the first detection environment feature set are obtained by surface collection and environmental feature collection, and K first detection images and a first detection environment feature set are obtained. The first detection environment feature set is then matched with the semantic information of the multiple detection scenes, and the first splicing and fusion routes are mapped according to the matching result to obtain the first splicing and fusion route, and then the K first detection images are fused by using the first splicing and fusion route to obtain the first detection fusion image, and finally the lightweight detection network layer is used to perform vehicle violation detection and identification on the first detection fusion image and the first regional fusion background corresponding to the first splicing and fusion route to obtain the vehicle violation detection result, thereby achieving the technical effect of improving the accuracy of vehicle violation detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the accompanying drawings of the embodiment of the present invention will be briefly introduced below. A flow chart is used in the present application to illustrate the operations performed by the system according to the embodiment of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps can be processed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or more operations can be removed from these processes.

[0023] Figure 1 A flow chart of a method for detecting traffic violations of an intelligent transportation vehicle provided in an embodiment of the present application.

[0024] Figure 2 A schematic diagram of the structure of a smart traffic vehicle violation detection system provided in an embodiment of the present application.

[0025] Explanation of the reference numerals: multiple receptive field fusion distribution field determination module 10, background picture splicing and fusion module 20, target detection area information acquisition module 30, first splicing and fusion route acquisition module 40, first detection fusion picture acquisition module 50, vehicle violation detection and recognition module 60. DETAILED DESCRIPTION

[0026] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0027] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0028] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms "first\second" involved are merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "including" and "having" and any variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by technicians in the technical field of this application. The terms used herein are for the purpose of describing the embodiments of the present application only.

[0029] The present application embodiment provides a method for detecting traffic violations of intelligent traffic vehicles, such as Figure 1 As shown, the method is applied to a smart traffic vehicle violation detection system, the smart traffic vehicle violation detection system is communicatively connected to a monitoring device array, and the method includes:

[0030] Step S100, the historical monitoring screens of the interactive monitoring device array are subjected to receptive field fusion identification to determine multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple detection scene semantic information and multiple splicing fusion routes, and the monitoring device array includes K monitoring devices, and the K monitoring devices include K device positioning identifiers. Specifically, historical monitoring screen data is collected from the monitoring device array, and the collected historical monitoring screens are preprocessed, including denoising, contrast enhancement and other operations to improve image quality. According to different features of the monitoring screen (such as light intensity, weather conditions, etc.), the monitoring screen is divided into multiple scenes, and each scene represents a specific monitoring environment. Semantic information is extracted for each scene to identify scene features, such as strong light, rain, night, etc. Based on the division of scenes and the extraction of semantic information, multiple receptive field fusion distribution fields are constructed. The receptive field fusion distribution field not only contains the semantic information of the detection scene, but also includes the splicing and fusion routes between different monitoring screens. The splicing and fusion routes refer to the paths or strategies adopted when fusing the same parts in the pictures taken by different monitoring devices. In different scenes, the splicing and fusion routes are different due to differences in image quality, lighting conditions, etc.

[0031] In a possible implementation, the historical monitoring screens of the interactive monitoring device array are subjected to receptive field fusion recognition to determine multiple receptive field fusion distribution fields, and step S100 further includes step S110, interactively performing simultaneous sequential screen collection of multiple historical time nodes on the K monitoring devices of the monitoring device array to obtain multiple historical monitoring screen sets, wherein each historical monitoring screen set includes K historical monitoring screens. Specifically, multiple historical time nodes are set, and these nodes represent different time periods or event points that need to be analyzed. The K monitoring devices in the interactive monitoring device array ensure that they simultaneously collect screens at the set time nodes. The screens collected by the K monitoring devices at each time node are collected to form multiple historical monitoring screen sets. Step S120, multiple historical detection scene semantic information of the multiple historical monitoring screen sets are collected, and semantic twin recognition is performed on the multiple historical detection scene semantic information to obtain multiple aggregated detection scene semantic information sets and multiple historical monitoring screen clusters. Specifically, the screens in each historical monitoring screen set are preliminarily analyzed to extract preliminary detection scene semantic information. Semantic twin recognition technology (a recognition technology based on semantic similarity, used to aggregate data with similar semantic features) is applied to aggregate similar scene semantic information to form multiple aggregated detection scene semantic information sets. Based on the aggregated scene semantic information, historical monitoring screens are grouped into multiple historical monitoring screen clusters, each of which contains a group of semantically similar monitoring screens.

[0032] Step S130, combining the K device location identifiers to perform receptive field fusion identification on the multiple historical monitoring screen clusters to obtain multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple splicing fusion routes. Specifically, the location identifier information of the K monitoring devices is used to determine the position and coverage of each monitoring device in the monitoring scene. Combining the location information of the monitoring device and the historical monitoring screen clusters, the monitoring screen is divided into receptive fields, that is, the specific area that each monitoring device can observe is determined. On the basis of the receptive field division, fusion identification is performed to build a more comprehensive monitoring scene model, and multiple receptive field fusion distribution fields are obtained after fusion, each receptive field fusion distribution field represents a complete monitoring area covered by multiple monitoring devices, and each receptive field fusion distribution field corresponds to a splicing fusion route, that is, the splicing fusion route between different monitoring screens. Step S140, feature fusion identification is performed on the multiple aggregated detection scene semantic information sets to obtain the multiple detection scene semantic information. Specifically, the multiple aggregated detection scene semantic information sets are further analyzed to extract key feature information. By applying feature fusion technology, feature information from different sets is fused to obtain more comprehensive semantic information of the detection scene. This implementation method achieves the technical effect of efficiently and accurately determining the fusion distribution field of multiple receptive fields by clustering similar surveillance images into different historical surveillance image clusters.

[0033] In a possible implementation, multiple historical detection scene semantic information of the multiple historical monitoring screen sets are collected, and semantic twin recognition is performed on the multiple historical detection scene semantic information to obtain multiple aggregated detection scene semantic information sets and multiple historical monitoring screen clusters. Step S120 further includes step S121, constructing a semantic twin recognition branch, wherein the semantic twin recognition branch is embedded with a twin network algorithm. Specifically, according to the principle of the twin network algorithm, a neural network branch is designed for semantic twin recognition. This branch contains two identical sub-networks, which share the same weights and structures, but process different input data respectively. The sub-networks judge the similarity between the input data by comparing their outputs (such as feature vectors). Step S122, randomly extract M historical detection scene semantic information from the multiple historical detection scene semantic information, and perform node authentication on the M historical detection scene semantic information. If the authentication is passed, the M historical detection scene semantic information is used as M recognition nodes, wherein M is a positive integer greater than or equal to 3. Specifically, M samples are selected from all historical detection scene semantic information using a random sampling method. The M randomly selected historical detection scene semantic information are authenticated, and the authentication ensures that these historical detection scene semantic information are valid, representative and different, and can be used as a reference point for subsequent comparison, and can be used for similarity comparison with other historical detection scene semantic information. The authentication process includes checking the integrity and accuracy of the data, and whether it meets the preset diversity requirements (for example, ensuring that the scene covers a variety of weather and time conditions such as sunny days, rainy days, cloudy days, and nights). If a historical detection scene semantic information passes the authentication, it is regarded as a valid identification node. Ultimately, the M historical detection scene semantic information that pass the authentication is set as M identification nodes. Step S123, respectively use the semantic twin recognition branch to twin the N historical detection scene semantic information and the M identification nodes, and divide the N historical detection scene semantic information into the identification node with the largest twin degree, and obtain M aggregated detection scene semantic information, wherein the sum of M and N is the total number of multiple historical detection scene semantic information. Specifically, the remaining N historical detection scene semantic information in the multiple historical detection scene semantic information are input into the semantic twin recognition branch one by one with the M recognition nodes. By comparing the similarity between each historical detection scene semantic information and the recognition node, these historical detection scene semantic information are divided into the aggregated detection scene semantic information to which the recognition node with the largest twin degree (similarity) belongs. In this way, each recognition node will eventually gather a group of historical detection scene semantic information similar to it. Step S124, use the M aggregated detection scene semantic information as multiple aggregated detection scene semantic information, and map to obtain multiple historical monitoring screen clusters.Specifically, after the aggregation of historical detection scene semantic information is completed, each aggregated detection scene semantic information set corresponds to a set of similar historical monitoring screens. Through the mapping relationship, these aggregated detection scene semantic information sets are mapped back to their corresponding original historical monitoring screens, thereby obtaining multiple historical monitoring screen clusters. Each cluster contains all monitoring screens belonging to the same aggregated detection scene semantic information. This implementation method significantly reduces the amount of calculation in the twin recognition process by randomly extracting and authenticating M historical detection scene semantic information as identification nodes, compared with using all historical detection scene semantic information as identification nodes, and the identification nodes are authenticated, so the similarity comparison results between the identification nodes are more accurate and reliable, thereby achieving the technical effect of improving the efficiency and accuracy of semantic twin recognition.

[0034] In a possible implementation, M historical detection scene semantic information is randomly extracted from the multiple historical detection scene semantic information, and node authentication is performed on the M historical detection scene semantic information. Step S122 further includes step S1221, enumerating and combining the M historical detection scene semantic information in pairs to obtain multiple historical detection scene semantic information combinations. Specifically, any two of the M historical detection scene semantic information are selected for combination to form a new combination. This process needs to be repeated until all possible combinations are enumerated. For example, if M=4, there will be C(4,2)=6 different combinations. Step S1222, using the semantic twin recognition branch to perform twin degree recognition on the multiple historical detection scene semantic information combinations to obtain multiple combined twin degrees. Specifically, each historical detection scene semantic information combination obtained in step S1221 is passed as input to the semantic twin recognition branch, and the semantic twin recognition branch outputs a numerical value indicating the similarity of the scene semantic information in the input combination, where 0 indicates complete dissimilarity and 1 indicates complete identity. The twin degree of each combination is recorded to form a set of multiple combined twin degrees. Step S1223, when the multiple combined twin degrees are all less than or equal to the preset twin degree threshold, the authentication is passed. Specifically, each combined twin degree obtained in step S1222 is compared with the preset twin degree threshold. If all combined twin degrees are less than or equal to the preset twin degree threshold, it means that there is enough difference between the M historical detection scene semantic information, which can be used as an identification node, so the authentication is passed. If there is any combined twin degree greater than the preset twin degree threshold, it means that there is too much similarity between the semantic information of these historical detection scenes. At this time, it is necessary to re-execute the random extraction and authentication process until the conditions are met. This implementation method ensures that the selected M historical detection scene semantic information are not only representative before being used as an identification node, but also have certain differences between them, avoiding the selection of nodes that are too similar or repeated, that is, avoiding the introduction of too much redundant or repeated information in the twin recognition process, thereby achieving the technical effect of improving the accuracy of semantic twin recognition.

[0035] In a possible implementation, the multiple historical monitoring screen clusters are subjected to receptive field fusion identification in combination with the K device positioning identifiers to obtain multiple receptive field fusion distribution fields, and step S130 further includes step S131, randomly extracting a first historical monitoring screen cluster from the multiple historical monitoring screen clusters, and then randomly extracting a first historical monitoring screen set from the first historical monitoring screen cluster. Specifically, the historical monitoring screen cluster is a set of historical monitoring screens grouped according to scenes, and each cluster represents a specific monitoring scenario. One is randomly selected from all historical monitoring screen clusters, and this cluster represents a set of monitoring screens at multiple time nodes in a specific scene. From the selected first historical monitoring screen cluster, a set of monitoring screens corresponding to a time node is further randomly selected, that is, a first historical monitoring screen set, which contains the screens collected by K monitoring devices at the same time node. Step S132, the first historical monitoring screen set is subjected to overlapping area identification in combination with the K device positioning identifiers to obtain multiple identified historical screen overlapping areas. Specifically, using the positioning identifiers (such as position coordinates, orientation, etc.) of K devices as references, the images in the first historical monitoring image set are spatially aligned, and the overlapping areas in these images are identified, that is, the areas commonly covered in the images of different monitoring devices (the intersection of multiple images), to form multiple identified historical image overlapping areas. Step S133, extracting the first image intersection set of the first identified historical image overlapping area of ​​the multiple identified historical image overlapping areas, taking any one of the first image intersections as the starting point, performing transition gradient recognition in the first identified historical image overlapping area according to a preset iteration step, and obtaining the first stage splicing and fusion point. Specifically, extract any one of the multiple identified historical image overlapping areas as the first identified historical image overlapping area, and then extract the first image intersection set from the first identified historical image overlapping area. Taking any one of the first image intersections as the starting point, performing transition gradient recognition in the first identified historical image overlapping area according to a preset iteration step and direction, and finding a point with a significant gradient change as the first stage splicing and fusion point. Step S134, taking the first stage splicing and fusion point as the stage starting point, multiple transition gradient identifications are performed to obtain the first splicing and fusion sub-route of the first identified historical screen overlapping area. Specifically, taking the first stage splicing and fusion point as the new starting point, the transition gradient identification is continued, and the path is gradually expanded until the entire first identified historical screen overlapping area is covered to form the first splicing and fusion sub-route.

[0036] Step S135, combining the K device positioning identifiers to perform receptive field fusion recognition on the overlapping areas of the multiple identified historical images, and obtain a first spliced ​​fusion sub-route set of the first historical monitoring image set; Step S136, combining the K device positioning identifiers to perform receptive field fusion recognition on the first historical monitoring image cluster, and obtain multiple spliced ​​fusion sub-route sets; Step S137, performing overlapping area fitting on the multiple spliced ​​fusion sub-route sets, and obtain a first spliced ​​fusion route of the first historical monitoring image cluster. Specifically, steps S133 to S134 are repeated for each overlapping area of ​​the identified historical images in the first historical monitoring image cluster to obtain multiple spliced ​​fusion sub-route sets. These sub-route sets are fitted with the overlapping area, that is, the sub-routes of the overlapping parts are fused and adjusted to obtain a more continuous and smooth first spliced ​​fusion route of the first historical monitoring image cluster. Step S138, using the first spliced ​​fusion route combined with the K monitoring ranges of the K monitoring devices to divide the target detection area, and obtain a first receptive field fusion distribution field, wherein the first receptive field fusion distribution field corresponds to a detection scene semantic information. Specifically, the target detection area is divided using the first splicing fusion route and the monitoring range of the K monitoring devices. The first receptive field fusion distribution field is constructed by integrating the position, orientation, monitoring range and splicing fusion route of the monitoring device, and the first receptive field fusion distribution field contains the semantic information of the detection scene. Step S139, combining the K device positioning identifiers to perform receptive field fusion identification on the multiple historical monitoring screen clusters to obtain multiple receptive field fusion distribution fields. Specifically, steps S131 to S138 are repeated for all historical monitoring screen clusters to construct a corresponding receptive field fusion distribution field for each historical monitoring screen cluster. Finally, multiple receptive field fusion distribution fields are obtained, each distribution field corresponding to a specific detection scene semantic information. This implementation method accurately determines the information density and change characteristics of the overlapping areas of different identified historical screens in the monitoring screen through transition gradient identification, provides an accurate starting point and path for receptive field fusion, and achieves the technical effect of improving the accuracy of receptive field fusion.

[0037] In a possible implementation, step S134 further includes step S1341, after multiple transition gradient identifications, until the screen intersection farthest from the starting point is reached in the first screen intersection set, identification is stopped, and multiple stage splicing and fusion points are obtained. Specifically, starting from the starting point, within a certain range around the current point (such as a 3x3 or 5x5 pixel area), the gradient value (the rate of change of grayscale value, used to indicate edge strength) of each pixel is calculated, and the point with the largest gradient value is found. This point is the transition point or splicing and fusion point of the current point, because it indicates that the image content has changed significantly. The current point is updated to the found gradient maximum value point, and the gradient calculation is continued around the new point. This process is repeated until the screen intersection farthest from the starting point is reached in the first screen intersection set. In this process, multiple transition points will be passed to obtain multiple stage splicing and fusion points. Step S1342, the first stage splicing and fusion point and the multiple stage splicing and fusion points are sequentially connected in the identification order to obtain the first splicing and fusion sub-route of the overlapping area of ​​the first identification history screen. Specifically, the splicing and fusion points (starting points) identified in the first stage and the splicing and fusion points of multiple stages obtained by subsequent gradient recognition are sorted in the recognition order. These points are connected in sequence using straight lines or curves (depending on the actual situation) to form a path from the starting point to the farthest intersection. This path is the first splicing and fusion sub-route of the overlapping area of ​​the first identified historical picture. This implementation method gradually approaches the farthest intersection through multiple transition gradient recognitions, ensuring the continuity and accuracy of the splicing path, and achieving the technical effect of improving the accuracy of image splicing path construction.

[0038] Step S200, stitching and fusing the background images of the target detection area according to the multiple stitching and fusion routes and the multiple receptive field fusion distribution fields to obtain multiple regional fusion backgrounds. Specifically, extract the background image from the monitoring device during the time period when no vehicle passes. According to the determined stitching and fusion routes, the background images under different monitoring devices are stitched to form multiple regional fusion backgrounds, that is, the regional fusion background refers to the pure background image in the target detection area after stitching and fusion under different detection scenarios.

[0039] Step S300, calling the K monitoring devices to collect images and environmental features of the target detection area within the preset detection window, and obtaining K first detection images and a first detection environmental feature set. Specifically, the preset detection window refers to the time period for detecting violations of vehicles in the target detection area. Within the preset detection window, calling the K monitoring devices to collect real-time images of the target detection area, and obtaining K first detection images. At the same time, the environmental feature data in the target detection area, such as light intensity, weather conditions, etc., are collected through sensors to form a first detection environmental feature set.

[0040] Step S400, using the first detection environment feature set to match the multiple detection scene semantic information, and mapping the multiple splicing and fusion routes based on the matching result to obtain a first splicing and fusion route. Specifically, the first detection environment feature set is matched with the extracted multiple detection scene semantic information to find the best matching scene. According to the mapping relationship between the detection scene semantic information and the splicing and fusion route, the first splicing and fusion route suitable for the current scene is obtained.

[0041] Step S500: fuse the K first detection images using the first splicing and fusion route to obtain a first detection fusion image. Specifically, fuse the K first detection images using the first splicing and fusion route to eliminate overlapping parts between the images to generate a first detection fusion image.

[0042] Step S600, using a lightweight detection network layer to perform vehicle violation detection and identification on the first detection fusion picture and the first regional fusion background corresponding to the first spliced ​​fusion route, and obtain a vehicle violation detection result. Specifically, the first detection fusion picture and the first regional fusion background of the corresponding scene are used as input and sent to the lightweight detection network layer. The lightweight detection network layer is a neural network model for fast and efficient processing of image data, with fewer parameters and calculations, and is used in application scenarios with high real-time requirements. The lightweight detection network layer processes the input data, identifies vehicle violations, such as running a red light, illegal lane changes, etc., and outputs vehicle violation detection results, including violation type, time, location and other information. The embodiment of the present application adopts technical means such as determining multiple receptive field fusion distribution fields through receptive field fusion recognition, performing background picture splicing and fusion and picture collection of the target detection area, and using a lightweight detection network layer to perform vehicle violation detection and identification on the fused picture, thereby achieving the technical effect of improving the accuracy of vehicle violation detection.

[0043] In the above, refer to Figure 1 A method for detecting traffic violations of a smart traffic vehicle according to an embodiment of the present invention is described in detail. Figure 2 An intelligent traffic vehicle violation detection system according to an embodiment of the present invention is described.

[0044] According to an embodiment of the present invention, a smart traffic vehicle violation detection system is used to solve the technical problem of insufficient detection accuracy of existing vehicle violation detection, and achieve the technical effect of improving the accuracy of vehicle violation detection. A smart traffic vehicle violation detection system includes: a plurality of receptive field fusion distribution field determination modules 10, a background image splicing and fusion module 20, a target detection area information acquisition module 30, a first splicing and fusion route acquisition module 40, a first detection fusion image acquisition module 50, and a vehicle violation detection and recognition module 60.

[0045] The multiple receptive field fusion distribution field determination module 10 is used to perform receptive field fusion identification on the historical monitoring images of the interactive monitoring device array, and determine multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple detection scene semantic information and multiple splicing fusion routes, and the monitoring device array includes K monitoring devices, and the K monitoring devices include K device positioning identifiers; the background image splicing and fusion module 20 is used to perform background image splicing and fusion of the target detection area according to the multiple splicing fusion routes and the multiple receptive field fusion distribution fields to obtain multiple regional fusion backgrounds; the target detection area information acquisition module 30 is used to call the K monitoring devices within a preset detection window to perform image acquisition and Environmental feature collection, obtaining K first detection images and a first detection environment feature set; a first splicing and fusion route acquisition module 40 is used to use the first detection environment feature set to match the multiple detection scene semantic information, and map the multiple splicing and fusion routes according to the matching results to obtain a first splicing and fusion route; a first detection fusion image acquisition module 50 is used to use the first splicing and fusion route to fuse the K first detection images to obtain a first detection fusion image; a vehicle violation detection and recognition module 60 is used to use a lightweight detection network layer to perform vehicle violation detection and recognition on the first detection fusion image and the first regional fusion background corresponding to the first splicing and fusion route to obtain a vehicle violation detection result.

[0046] The specific configuration of the multiple receptive field fusion distribution field determination module 10 will be described in detail below. As described above, the historical monitoring screens of the interactive monitoring device array are subjected to receptive field fusion recognition to determine multiple receptive field fusion distribution fields. The multiple receptive field fusion distribution field determination module 10 may further include: a simultaneous sequence picture acquisition unit is used to interact with the K monitoring devices of the monitoring device array, perform simultaneous sequence picture acquisition of multiple historical time nodes, and obtain multiple historical monitoring picture sets, wherein each historical monitoring picture set includes K historical monitoring pictures; a semantic twin recognition unit is used to collect multiple historical detection scene semantic information of the multiple historical monitoring picture sets, perform semantic twin recognition on the multiple historical detection scene semantic information, and obtain multiple aggregated detection scene semantic information sets and multiple historical monitoring picture clusters; a receptive field fusion recognition unit is used to perform receptive field fusion recognition on the multiple historical monitoring picture clusters in combination with the K device positioning identifiers, and obtain multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple splicing fusion routes; a feature fusion recognition unit is used to perform feature fusion recognition on the multiple aggregated detection scene semantic information sets to obtain the multiple detection scene semantic information.

[0047] Among them, multiple historical detection scene semantic information of the multiple historical monitoring screen sets are collected, and semantic twin recognition is performed on the multiple historical detection scene semantic information to obtain multiple aggregated detection scene semantic information sets and multiple historical monitoring screen clusters. The semantic twin recognition unit may further include: a semantic twin recognition branch construction subunit for constructing a semantic twin recognition branch, wherein the semantic twin recognition branch is embedded with a twin network algorithm; a node authentication subunit is used to randomly extract M historical detection scene semantic information from the multiple historical detection scene semantic information, and perform node authentication on the M historical detection scene semantic information. If the authentication is passed, then The M historical detection scene semantic information is used as M recognition nodes, where M is a positive integer greater than or equal to 3; the twin recognition subunit is used to use the semantic twin recognition branch to perform twin recognition on the N historical detection scene semantic information and the M recognition nodes, respectively, and divide the N historical detection scene semantic information to the recognition node with the largest twin degree to obtain M aggregated detection scene semantic information, where the sum of M and N is the total number of multiple historical detection scene semantic information; the multiple historical monitoring screen cluster acquisition subunit is used to use the M aggregated detection scene semantic information as multiple aggregated detection scene semantic information, and map to obtain multiple historical monitoring screen clusters.

[0048] Among them, M historical detection scene semantic information are randomly extracted from the multiple historical detection scene semantic information, and node authentication is performed on the M historical detection scene semantic information. The node authentication subunit may further include: a pairwise enumeration combination micro-unit is used to enumerate and combine the M historical detection scene semantic information in pairs to obtain multiple historical detection scene semantic information combinations; a twin degree identification micro-unit is used to use the semantic twin identification branch to perform twin degree identification on the multiple historical detection scene semantic information combinations to obtain multiple combined twin degrees; an authentication micro-unit is used to authenticate when the multiple combined twin degrees are all less than or equal to a preset twin degree threshold.

[0049] Among them, the receptive field fusion recognition is performed on the multiple historical monitoring screen clusters in combination with the K device positioning identifiers to obtain multiple receptive field fusion distribution fields. The receptive field fusion recognition unit may further include: a random extraction subunit is used to randomly extract a first historical monitoring screen cluster from the multiple historical monitoring screen clusters, and then randomly extract a first historical monitoring screen set from the first historical monitoring screen cluster; an overlapping area recognition subunit is used to perform overlapping area recognition on the first historical monitoring screen set in combination with the K device positioning identifiers to obtain multiple identified historical screen overlapping areas; a first-stage splicing and fusion point acquisition subunit is used to extract a first screen intersection set of the first identified historical screen overlapping area of ​​the multiple identified historical screen overlapping areas, and take any one of the first screen intersection points in the first screen intersection point set as the starting point, perform transition gradient recognition in the first identified historical screen overlapping area according to a preset iteration step size, and obtain a first-stage splicing and fusion point; a first splicing and fusion sub-route acquisition subunit is used to take the first-stage splicing and fusion point as the stage starting point, perform multiple transition gradient recognitions, and obtain a first identified historical screen overlapping area. splicing fusion sub-routes; a first splicing fusion sub-route set acquisition sub-unit is used to perform receptive field fusion identification on the overlapping areas of the multiple identified historical images in combination with the K device positioning identifiers to obtain a first splicing fusion sub-route set of the first historical monitoring image set; multiple first splicing fusion sub-route set acquisition sub-units are used to perform receptive field fusion identification on the first historical monitoring image cluster in combination with the K device positioning identifiers to obtain multiple splicing fusion sub-route sets; a same overlapping area fitting sub-unit is used to perform same overlapping area fitting on the multiple splicing fusion sub-route sets to obtain a first splicing fusion route of the first historical monitoring image cluster; a first receptive field fusion distribution field acquisition sub-unit is used to divide the target detection area by using the first splicing fusion route in combination with the K monitoring ranges of the K monitoring devices to obtain a first receptive field fusion distribution field, wherein the first receptive field fusion distribution field corresponds to a detection scene semantic information; multiple receptive field fusion distribution field acquisition sub-units are used to perform receptive field fusion identification on the multiple historical monitoring image clusters in combination with the K device positioning identifiers to obtain multiple receptive field fusion distribution fields.

[0050] Among them, the first splicing and fusion sub-route acquisition sub-unit may further include: a multi-stage splicing and fusion point acquisition micro-unit is used to perform multiple transition gradient identifications until the picture intersection farthest from the starting point in the first picture intersection set is reached, and then the identification is stopped to obtain multiple stage splicing and fusion points; the first splicing and fusion sub-route acquisition micro-unit is used to connect the first stage splicing and fusion points and the multiple stage splicing and fusion points in sequence according to the identification order to obtain the first splicing and fusion sub-route of the overlapping area of ​​the first identified historical pictures.

[0051] An intelligent traffic vehicle violation detection system provided by an embodiment of the present invention can execute an intelligent traffic vehicle violation detection method provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0052] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0053] The above specific implementation manner does not constitute a limitation to the protection scope of the present application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application. In some cases, the actions or steps recorded in the present application can be performed in an order different from that in the embodiment and can still achieve the desired results. In addition, the process depicted in the accompanying drawings does not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for detecting traffic violations of intelligent traffic vehicles, characterized in that: The method is applied to a smart traffic vehicle violation detection system, wherein the smart traffic vehicle violation detection system is communicatively connected to a monitoring device array, and the method comprises: Perform receptive field fusion recognition on historical monitoring images of an interactive monitoring device array to determine multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple detection scene semantic information and multiple splicing fusion routes, and the monitoring device array includes K monitoring devices, and the K monitoring devices include K device positioning identifiers; Performing stitching and fusion of background images of the target detection area according to the multiple stitching and fusion routes and the multiple receptive field fusion distribution fields to obtain multiple regional fusion backgrounds; Calling the K monitoring devices to collect images and environmental features of the target detection area within a preset detection window to obtain K first detection images and a first detection environmental feature set; Using the first detection environment feature set to match the plurality of detection scene semantic information, and mapping the plurality of splicing and fusion routes according to the matching result to obtain a first splicing and fusion route; Using the first splicing and fusion route to fuse the K first detection images to obtain a first detection fusion image; Using a lightweight detection network layer, performing vehicle violation detection and recognition on the first detection fusion picture and the first regional fusion background corresponding to the first spliced ​​fusion route to obtain a vehicle violation detection result; The historical monitoring images of the interactive monitoring device array are used for receptive field fusion recognition to determine multiple receptive field fusion distribution fields, including: Interacting K monitoring devices of the monitoring device array to collect simultaneous images of multiple historical time nodes to obtain multiple historical monitoring image sets, wherein each historical monitoring image set includes K historical monitoring images; Collecting multiple historical detection scene semantic information of the multiple historical monitoring screen sets, performing semantic twin recognition on the multiple historical detection scene semantic information, and obtaining multiple aggregated detection scene semantic information sets and multiple historical monitoring screen clusters; Combining the K device positioning identifiers, performing receptive field fusion identification on the multiple historical monitoring screen clusters to obtain multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple splicing fusion routes; Perform feature fusion recognition on the multiple aggregated detection scene semantic information sets to obtain the multiple detection scene semantic information.

2. A method for detecting traffic violations of intelligent traffic vehicles as claimed in claim 1, characterized in that: Collecting multiple historical detection scene semantic information of the multiple historical monitoring screen sets, performing semantic twin recognition on the multiple historical detection scene semantic information, and obtaining multiple aggregated detection scene semantic information sets and multiple historical monitoring screen clusters, including: Constructing a semantic twin recognition branch, wherein the semantic twin recognition branch is embedded with a twin network algorithm; Randomly extract M historical detection scene semantic information from the multiple historical detection scene semantic information, and perform node authentication on the M historical detection scene semantic information. If the authentication is passed, the M historical detection scene semantic information is used as M identification nodes, where M is a positive integer greater than or equal to 3; Using the semantic twin recognition branches respectively, the N historical detection scene semantic information and the M recognition nodes are twin recognized, and the N historical detection scene semantic information are divided into the recognition nodes with the largest twin degree, to obtain M aggregated detection scene semantic information, wherein the sum of M and N is the total number of multiple historical detection scene semantic information; The M aggregated detection scene semantic information are used as multiple aggregated detection scene semantic information, and mapped to obtain multiple historical monitoring screen clusters.

3. A method for detecting traffic violations of intelligent traffic vehicles as claimed in claim 2, characterized in that: Randomly extracting M pieces of historical detection scene semantic information from the plurality of historical detection scene semantic information, and performing node authentication on the M pieces of historical detection scene semantic information, including: Enumerate and combine the M historical detection scene semantic information in pairs to obtain multiple historical detection scene semantic information combinations; Using the semantic twin recognition branch, the multiple historical detection scene semantic information combinations are subjected to twin degree recognition to obtain multiple combined twin degrees; When the multiple combined twin degrees are all less than or equal to the preset twin degree threshold, the authentication is passed.

4. A method for detecting traffic violations of intelligent traffic vehicles as claimed in claim 1, characterized in that: Combine the K device location identifiers to perform receptive field fusion identification on the multiple historical monitoring screen clusters to obtain multiple receptive field fusion distribution fields, including: Randomly extracting a first historical monitoring picture cluster from the plurality of historical monitoring picture clusters, and then randomly extracting a first historical monitoring picture set from the first historical monitoring picture cluster; identifying overlapping areas of the first historical monitoring screen set in combination with the K device positioning identifiers to obtain a plurality of identified historical screen overlapping areas; Extracting a first screen intersection point set of a first identified historical screen overlap area of ​​the plurality of identified historical screen overlap areas, taking any screen intersection point in the first screen intersection point set as a starting point, performing transition gradient identification in the first identified historical screen overlap area according to a preset iteration step length, and obtaining a first-stage splicing fusion point; Then, taking the first stage splicing and fusion point as the stage starting point, multiple transition gradient identifications are performed to obtain the first splicing and fusion sub-route of the overlapping area of ​​the first identified historical images; Based on the multiple historical monitoring screen clusters, receptive field fusion identification is performed to obtain multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple splicing fusion routes.

5. A method for detecting traffic violations of intelligent traffic vehicles as claimed in claim 4, characterized in that: include: Combining the K device positioning identifiers, performing receptive field fusion recognition on the overlapping areas of the multiple identified historical images, and obtaining a first spliced ​​fusion sub-route set of the first historical monitoring image set; Combining the K device location identifiers, performing receptive field fusion recognition on the first historical monitoring screen cluster to obtain multiple spliced ​​fusion sub-route sets; Performing overlapping region fitting on the plurality of splicing and fusion sub-route sets to obtain a first splicing and fusion route of the first historical monitoring screen cluster; The target detection area is divided by using the first splicing and fusion route in combination with the K monitoring ranges of the K monitoring devices to obtain a first receptive field fusion distribution field, wherein the first receptive field fusion distribution field corresponds to a detection scene semantic information; The K device positioning identifiers are combined to perform receptive field fusion identification on the multiple historical monitoring image clusters to obtain multiple receptive field fusion distribution fields.

6. A method for detecting traffic violations of intelligent traffic vehicles as claimed in claim 4, characterized in that: include: After multiple transition gradient identifications, when the image intersection point farthest from the starting point in the first image intersection point set is reached, identification is stopped to obtain multiple stage splicing and fusion points; The first-stage splicing and fusion points and the multiple-stage splicing and fusion points are sequentially connected in the recognition order to obtain a first splicing and fusion sub-route of the first recognized historical screen overlapping area.

7. A smart traffic vehicle violation detection system, characterized in that: The system is used to implement a smart traffic vehicle violation detection method according to any one of claims 1 to 6, and the system comprises: A module for determining a plurality of receptive field fusion distribution fields, wherein the module is used to perform receptive field fusion identification on historical monitoring images of an interactive monitoring device array, and determine a plurality of receptive field fusion distribution fields, wherein the plurality of receptive field fusion distribution fields include a plurality of detection scene semantic information and a plurality of splicing fusion routes, and the monitoring device array includes K monitoring devices, and the K monitoring devices include K device positioning identifiers; A background image stitching and fusion module, the background image stitching and fusion module is used to stitch and fuse the background images of the target detection area according to the multiple stitching and fusion routes and the multiple receptive field fusion distribution fields to obtain multiple regional fusion backgrounds; A target detection area information acquisition module, the target detection area information acquisition module is used to call the K monitoring devices to collect images and environmental features of the target detection area within a preset detection window, and obtain K first detection images and a first detection environment feature set; A first splicing and fusion route acquisition module, the first splicing and fusion route acquisition module is used to match the first detection environment feature set with the plurality of detection scene semantic information, and map the plurality of splicing and fusion routes according to the matching result to acquire a first splicing and fusion route; A first detection fusion picture acquisition module, the first detection fusion picture acquisition module is used to fuse the K first detection pictures by using the first splicing and fusion route to obtain a first detection fusion picture; A vehicle violation detection and identification module, the vehicle violation detection and identification module is used to use a lightweight detection network layer to perform vehicle violation detection and identification on the first detection fusion picture and the first regional fusion background corresponding to the first spliced ​​fusion route to obtain a vehicle violation detection result; The module for determining multiple receptive field fusion distribution fields further includes: a simultaneous sequential picture acquisition unit for interacting with the K monitoring devices of the monitoring device array, performing simultaneous sequential picture acquisition at multiple historical time nodes, and obtaining multiple historical monitoring picture sets, wherein each historical monitoring picture set includes K historical monitoring pictures; a semantic twin recognition unit for collecting multiple historical detection scene semantic information of the multiple historical monitoring picture sets, performing semantic twin recognition on the multiple historical detection scene semantic information, and obtaining multiple aggregated detection scene semantic information sets and multiple historical monitoring picture clusters; a receptive field fusion recognition unit for performing receptive field fusion recognition on the multiple historical monitoring picture clusters in combination with the K device positioning identifiers, and obtaining multiple receptive field fusion distribution fields, wherein the multiple receptive field fusion distribution fields include multiple splicing fusion routes; a feature fusion recognition unit for performing feature fusion recognition on the multiple aggregated detection scene semantic information sets, and obtaining the multiple detection scene semantic information.

Images