Security monitoring video artificial intelligence identification method and system based on adaptive scene matching

Abstract

The invention relates to the technical field of artificial intelligence, and provides a self-adaptive scene matching safety monitoring video artificial intelligence identification method and system, and the method comprises the steps: detecting object motion state data in a target site, determining scene attribute information, and adjusting a video monitoring operation state of the target site; based on the feature information of the video monitoring operation state and the first picture feature of the monitoring video, the monitoring video is divided and classified to obtain a plurality of monitoring sub-video sets, each monitoring sub-video set is input into a matched neural network model for recognition, and a video picture recognition result is obtained; different monitoring sub-videos are subjected to matched artificial intelligence identification, so that effective monitoring sub-videos in the monitoring sub-videos are screened and reserved; the abnormal event occurrence attribute information of the target site is determined based on the respective video picture recognition results of all the effective monitoring sub-videos, and the monitoring video is subjected to identification processing, so that scene matching processing of the target site is realized, and the security recognition analysis flexibility and accuracy of the monitoring video are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to an AI-based method and system for adaptive scene matching in security surveillance video recognition. Background Technology

[0002] Video surveillance is widely used in public place security. By capturing and analyzing real-time video footage of public places, potential dangers can be detected promptly. Furthermore, to improve the efficiency and accuracy of video surveillance analysis, neural network models are used to process the video, effectively and quickly identifying and predicting safety hazards in public places. Current video surveillance analysis relies on single-type neural network models. However, the image quality of surveillance videos is not always stable during actual capture, leading to varying analysis requirements. Using a single-type neural network model for comprehensive analysis throughout the entire process cannot guarantee optimal processing, reducing the flexibility of analysis and preventing scene-specific AI-based recognition, thus lowering accuracy. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an adaptive scene-matching AI-powered identification method and system for security surveillance videos. It detects the motion state data of objects within the target site, determines scene attribute information, and adjusts the video surveillance operation state accordingly to improve the matching accuracy of the target site. Based on the feature information of the video surveillance operation state and the first-frame features of the surveillance video, the surveillance video is divided and categorized into several sets of sub-videos. Each sub-video set is input into a matching neural network model for identification, resulting in video image recognition results. This allows different sub-videos to be matched and identified using AI, thereby filtering and retaining the effective sub-videos. Based on the video image recognition results of all effective sub-videos, the abnormal event occurrence attribute information of the target site is determined. The surveillance video is then further processed and labeled to achieve scene-based matching of the target site, improving the flexibility and accuracy of security identification and analysis of surveillance videos, and ensuring the overall efficiency and reliability of the surveillance video.

[0004] This invention provides an AI-powered method for adaptive scene matching in security surveillance video recognition, comprising the following steps:

[0005] Step S1: Detect the target scene to obtain object motion state data within the target scene; based on the object motion state data, obtain scene attribute information of the target scene; based on the scene attribute information, adjust the video surveillance operation status of the target scene.

[0006] Step S2: Based on the feature information of the video monitoring operation status, the monitoring video is divided into several monitoring sub-videos; frame sampling analysis is performed on all monitoring sub-videos to determine the first frame feature information of all monitoring sub-videos, thereby dividing all monitoring sub-videos into several monitoring sub-video sets.

[0007] Step S3: Based on the second screen feature information of each of the monitoring sub-video sets, input each monitoring sub-video set into the matching neural network model for recognition to obtain the corresponding video screen recognition result; based on the video screen recognition result, filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos;

[0008] Step S4: Based on the video image recognition results of each of the valid monitoring sub-videos, determine the abnormal event occurrence attribute information of the target site; based on the abnormal event occurrence attribute information, perform identification processing on the monitoring video.

[0009] In one embodiment disclosed in this application, in step S1, the target scene is detected to obtain object motion state data within the target scene; based on the object motion state data, scene attribute information of the target scene is obtained; based on the scene attribute information, the video surveillance operation state of the target scene is adjusted, including:

[0010] Non-contact scanning detection is performed on the target site to obtain the motion trajectory data and motion speed data of all objects inside the target site, which are used as the motion state data of the objects.

[0011] The motion trajectory data and motion speed data are analyzed to determine the object occurrence probability information and object distribution density information for each of the sub-regions under the target site, which are used as the scene attribute information. Based on the object occurrence probability information and the object distribution density information, the shooting operation parameters corresponding to video monitoring of each sub-region under the target site are adjusted. The shooting operation parameters include at least one of the shooting field of view and the shooting focal length.

[0012] In one embodiment of this application, in step S2, based on the feature information of the video monitoring operation state, the monitoring video is divided into several monitoring sub-videos; frame sampling analysis is performed on each monitoring sub-video to determine the first frame feature information of each monitoring sub-video, thereby dividing all monitoring sub-videos into several monitoring sub-video sets, including:

[0013] Based on the time characteristic information of the shooting operation parameters change during the adjustment of the video monitoring operation status, the entire video monitoring process of the target site is divided into several shooting time intervals; based on the start and end points of all shooting time intervals, the monitoring video is divided into several monitoring sub-videos.

[0014] For each monitoring sub-video, frames are randomly extracted, and pixel texture recognition processing is performed on several randomly extracted frames to obtain the pixel texture feature information corresponding to each monitoring sub-video, which is used as the first image feature information. Based on the similarity of the pixel texture feature information between all monitoring sub-videos, all monitoring sub-videos are divided into several monitoring sub-video sets. Among them, the similarity of the pixel texture feature information between all monitoring sub-videos under each monitoring sub-video set satisfies the preset similarity distribution condition.

[0015] In one embodiment disclosed in this application, in step S3, based on the second-screen feature information of each of the monitoring sub-video sets, each monitoring sub-video set is input into a matching neural network model for recognition to obtain the corresponding video screen recognition result; based on the video screen recognition result, all monitoring sub-videos are filtered to retain the corresponding valid monitoring sub-videos, including:

[0016] For each monitoring sub-video set, pixel contour recognition processing is performed on all monitoring sub-videos to obtain the pixel contour distribution information of each monitoring sub-video, which is used as the second image feature information. The pixel contour distribution information is compared with the image type information of each neural network model in the neural network model library, so that each monitoring sub-video set is input into the matching neural network model for recognition, thereby obtaining the corresponding video image recognition result.

[0017] The video image recognition results are analyzed for object recognition credibility to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility is then compared with a threshold to filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos.

[0018] In one embodiment disclosed in this application, in step S4, based on the video frame recognition results of each of the valid monitoring sub-videos, the abnormal event occurrence attribute information of the target site is determined; based on the abnormal event occurrence attribute information, the monitoring video is processed for identification, including:

[0019] The video image recognition results of each of the valid monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding objects inside the target site; wherein, the action behavior feature information includes the type of action issued by the object, the range of influence of the action, and the time of the action;

[0020] Based on the type of action and the scope of influence of the action issued by the object, it is determined whether an abnormal event has occurred at the target site; if so, the corresponding action occurrence time information is used as the abnormal event occurrence attribute information.

[0021] Based on the action occurrence time information corresponding to the abnormal event occurrence attribute information, the timeline of the monitoring video is marked with time points, thereby marking and processing the monitoring video for all abnormal event occurrence time points.

[0022] This invention also provides an AI-powered security surveillance video recognition system with adaptive scene matching, comprising:

[0023] The target scene recognition module is used to detect the target scene and obtain the motion state data of objects within the target scene; based on the object motion state data, the scene attribute information of the target scene is obtained.

[0024] The video surveillance operation adjustment module is used to adjust the video surveillance operation status of the target site based on the scene attribute information.

[0025] The surveillance video segmentation module is used to divide the surveillance video into several surveillance sub-videos based on the feature information of the video surveillance operation status.

[0026] The surveillance video integration module is used to perform frame sampling analysis on all surveillance sub-videos to determine the first frame feature information of all surveillance sub-videos, thereby dividing all surveillance sub-videos into several surveillance sub-video sets.

[0027] The artificial intelligence recognition module is used to input each monitoring sub-video set into a matching neural network model for recognition based on the second-screen feature information of each of the monitoring sub-video sets, and obtain the corresponding video screen recognition results.

[0028] The surveillance video filtering module is used to filter all surveillance sub-videos based on the video image recognition results and retain the corresponding valid surveillance sub-videos.

[0029] The abnormal event attribute determination module is used to determine the abnormal event occurrence attribute information of the target site based on the video image recognition results of each of the valid monitoring sub-videos.

[0030] The surveillance video identification module is used to identify and process the surveillance video based on the abnormal event occurrence attribute information.

[0031] In one embodiment disclosed in this application, the target scene recognition module is used to detect the target scene and obtain object motion state data within the target scene; based on the object motion state data, it obtains scene attribute information of the target scene, including:

[0032] Non-contact scanning detection is performed on the target site to obtain the motion trajectory data and motion speed data of all objects inside the target site, which are used as the motion state data of the objects.

[0033] The motion trajectory data and the motion speed data are analyzed to determine the object occurrence probability information and object distribution density information corresponding to each of the sub-regions under the target site, which are then used as the scene attribute information.

[0034] The video surveillance operation adjustment module is used to adjust the video surveillance operation status of the target site based on the scene attribute information, including:

[0035] Based on the object occurrence probability information and the object distribution density information, the shooting operation parameters corresponding to video monitoring of each sub-area under the target site are adjusted; wherein, the shooting operation parameters include at least one of the shooting field of view range and the shooting focal length.

[0036] In one embodiment disclosed in this application, the surveillance video segmentation module is used to divide the surveillance video into several surveillance sub-videos based on the feature information of the video surveillance operation status, including:

[0037] Based on the time characteristic information of the shooting operation parameters change during the adjustment of the video monitoring operation status, the entire video monitoring process of the target site is divided into several shooting time intervals; based on the start and end points of all shooting time intervals, the monitoring video is divided into several monitoring sub-videos.

[0038] The surveillance video integration module is used to perform frame sampling analysis on all surveillance sub-videos to determine the first-frame feature information of each surveillance sub-video, thereby dividing all surveillance sub-videos into several surveillance sub-video sets, including:

[0039] For each monitoring sub-video, frames are randomly extracted, and pixel texture recognition processing is performed on several randomly extracted frames to obtain the pixel texture feature information corresponding to each monitoring sub-video, which is used as the first image feature information. Based on the similarity of the pixel texture feature information between all monitoring sub-videos, all monitoring sub-videos are divided into several monitoring sub-video sets. Among them, the similarity of the pixel texture feature information between all monitoring sub-videos under each monitoring sub-video set satisfies the preset similarity distribution condition.

[0040] In one embodiment disclosed in this application, the artificial intelligence recognition module is used to input each monitoring sub-video set into a matching neural network model for recognition based on the second-screen feature information of each of the monitoring sub-video sets, to obtain the corresponding video screen recognition result, including:

[0041] For each monitoring sub-video set, pixel contour recognition processing is performed on all monitoring sub-videos to obtain the pixel contour distribution information of each monitoring sub-video, which is used as the second image feature information. The pixel contour distribution information is compared with the image type information of each neural network model in the neural network model library, so that each monitoring sub-video set is input into the matching neural network model for recognition, thereby obtaining the corresponding video image recognition result.

[0042] The surveillance video filtering module is used to filter all surveillance sub-videos based on the video image recognition results, retaining the corresponding valid surveillance sub-videos, including:

[0043] The video image recognition results are analyzed for object recognition credibility to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility is then compared with a threshold to filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos.

[0044] In one embodiment disclosed in this application, the abnormal event attribute determination module is used to determine the abnormal event occurrence attribute information of the target site based on the video frame recognition results of each of the valid monitoring sub-videos, including:

[0045] The video image recognition results of each of the valid monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding objects inside the target site; wherein, the action behavior feature information includes the type of action issued by the object, the range of influence of the action, and the time of the action;

[0046] Based on the type of action and the scope of influence of the action issued by the object, it is determined whether an abnormal event has occurred at the target site; if so, the corresponding action occurrence time information is used as the abnormal event occurrence attribute information.

[0047] The surveillance video identification module is used to identify and process the surveillance video based on the abnormal event occurrence attribute information, including:

[0048] Based on the action occurrence time information corresponding to the abnormal event occurrence attribute information, the timeline of the monitoring video is marked with time points, thereby marking and processing the monitoring video for all abnormal event occurrence time points.

[0049] Compared to existing technologies, this adaptive scene-matching AI-powered security monitoring video recognition method and system detects the motion state data of objects within the target site, determines scene attribute information, and adjusts the video monitoring operation state of the target site accordingly, thereby improving the targeting accuracy of the video capture. Based on the feature information of the video monitoring operation state and the first-frame features of the monitoring video, the monitoring video is divided and categorized into several monitoring sub-video sets. Each monitoring sub-video set is input into a matching neural network model for recognition, resulting in video image recognition results. This allows different monitoring sub-videos to be matched and recognized by AI, thereby filtering and retaining the effective monitoring sub-videos. Based on the video image recognition results of all effective monitoring sub-videos, the abnormal event occurrence attribute information of the target site is determined, and the monitoring video is then labeled and processed to achieve scene-based matching processing of the target site. This improves the flexibility and accuracy of security identification and analysis of monitoring videos, ensuring the overall efficiency and reliability of the monitoring video.

[0050] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1A flowchart illustrating the adaptive scene matching AI-powered recognition method for security surveillance videos provided by this invention;

[0054] Figure 2 This is a schematic diagram of the framework of the adaptive scene matching security monitoring video artificial intelligence recognition system provided by the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] See Figure 1 This is a flowchart illustrating the adaptive scene matching AI-powered recognition method for security surveillance videos provided in an embodiment of the present invention. The adaptive scene matching AI-powered recognition method for security surveillance videos includes:

[0057] Step S1: Detect the target scene to obtain the motion state data of objects within the target scene; based on the object motion state data, obtain the scene attribute information of the target scene; based on the scene attribute information, adjust the video surveillance operation status of the target scene.

[0058] Step S2: Based on the feature information of the video monitoring operation status, the monitoring video is divided into several monitoring sub-videos; frame sampling analysis is performed on all monitoring sub-videos to determine the first frame feature information of all monitoring sub-videos, thereby dividing all monitoring sub-videos into several monitoring sub-video sets.

[0059] Step S3: Based on the second-screen feature information of each of the monitoring sub-video sets, input each monitoring sub-video set into the matching neural network model for recognition to obtain the corresponding video screen recognition result; based on the video screen recognition result, filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos.

[0060] Step S4: Based on the video image recognition results of each of the valid monitoring sub-videos, determine the abnormal event occurrence attribute information of the target site; based on the abnormal event occurrence attribute information, perform identification processing on the monitoring video.

[0061] This adaptive scene-matching AI-powered security monitoring video recognition method detects the motion state data of objects within the target site, determines scene attribute information, and adjusts the video monitoring operation state of the target site accordingly, thereby improving the matching accuracy of the footage. Based on the feature information of the video monitoring operation state and the first-frame features of the monitoring video, the monitoring video is divided and categorized into several monitoring sub-video sets. Each monitoring sub-video set is input into a matching neural network model for recognition, resulting in video image recognition results. This allows different monitoring sub-videos to be matched and recognized by AI, thereby filtering and retaining the effective monitoring sub-videos. Based on the video image recognition results of all effective monitoring sub-videos, the abnormal event occurrence attribute information of the target site is determined. The monitoring video is then further processed and labeled to achieve scene-based matching of the target site, improving the flexibility and accuracy of security identification and analysis of monitoring videos, and ensuring the overall efficiency and reliability of the monitoring video.

[0062] Preferably, in step S1, the target scene is detected to obtain object motion state data within the target scene; based on the object motion state data, scene attribute information of the target scene is obtained; based on the scene attribute information, the video surveillance operation status of the target scene is adjusted, including:

[0063] Non-contact scanning detection is performed on the target site to obtain the motion trajectory data and motion speed data of all objects inside the target site, which are used as the motion state data of the objects.

[0064] The motion trajectory data and motion speed data are analyzed to determine the object occurrence probability information and object distribution density information for each sub-area under the target site, which are used as the scene attribute information. Based on the object occurrence probability information and object distribution density information, the shooting operation parameters corresponding to video monitoring of each sub-area under the target site are adjusted. The shooting operation parameters include at least one of the shooting field of view and shooting focal length.

[0065] In the aforementioned technical solutions, the flow of people or movement of vehicles at target sites such as public places are complex and variable. The movement and distribution of objects (people and / or objects) in different areas of the target site vary at different times. If a uniform video shooting method is always used when monitoring the target site, it is impossible to guarantee detailed global monitoring of different areas within the target site, and it is impossible to accurately identify the dynamic changes of objects within each area. Therefore, non-contact scanning detection, such as radar scanning or structured light scanning, is first performed on the target site to detect the movement trajectory and speed of all objects within the target site during the corresponding time period. This obtains the movement trajectory data and speed data of each object, thereby quantitatively determining the movement of objects in different areas of the target site. Then, statistical analysis is performed on the motion trajectory data and motion speed data to obtain the object occurrence probability information and object distribution density information for each sub-region under the target site. These sub-regions can be, but are not limited to, pre-divided grid sub-regions within the target site. The statistical calculation of the object occurrence probability and object distribution density (i.e., the number of objects per unit area) is a conventional technique in this field and will not be detailed here. Based on this object occurrence probability and distribution density information, the shooting field of view and / or shooting focal length for video surveillance of each sub-region under the target site are adjusted. For example, if the object occurrence probability within a sub-region exceeds a preset probability threshold, the shooting focal length for that sub-region is increased to achieve accurate shooting of objects within that sub-region; if the object distribution density within a sub-region exceeds a preset density threshold, the shooting field of view for that sub-region is increased so that all objects appearing within that sub-region can appear within the corresponding shooting field of view. This allows for global quantitative identification of the object presence within the target site, enabling targeted identification of different objects for subsequent assessment of the safety status within the scene.

[0066] Preferably, in step S2, based on the feature information of the video monitoring operation status, the monitoring video is divided into several monitoring sub-videos; frame sampling analysis is performed on each monitoring sub-video to determine the first frame feature information of each monitoring sub-video, thereby dividing all monitoring sub-videos into several monitoring sub-video sets, including:

[0067] Based on the time characteristics of the shooting operation parameters change during the adjustment of the video surveillance operation status, the entire video surveillance process of the target site is divided into several shooting time intervals; based on the start and end points of all shooting time intervals, the surveillance video is divided into several surveillance sub-videos.

[0068] For each monitoring sub-video, frames are randomly extracted, and pixel texture recognition processing is performed on several randomly extracted frames to obtain the pixel texture feature information corresponding to each monitoring sub-video, which is used as the first image feature information. Based on the similarity of the pixel texture feature information between all monitoring sub-videos, all monitoring sub-videos are divided into several monitoring sub-video sets. Among them, the similarity of the pixel texture feature information between all monitoring sub-videos under each monitoring sub-video set satisfies the preset similarity distribution condition.

[0069] In the above technical solution, during the video surveillance shooting of the target site, the corresponding shooting field of view and / or shooting focal length are adjusted so that these ranges change over time. Based on the characteristic information of these changes, the entire video surveillance process of the target site is divided into several shooting time intervals. Within each shooting time interval, the fluctuation rates of the shooting field of view and shooting focal length remain within their respective ranges, ensuring that the surveillance video footage obtained for each shooting time interval is relatively stable. Furthermore, based on the start and end points of all shooting time intervals, the surveillance video is divided into several monitoring sub-videos, each corresponding to a shooting time interval, thus achieving effective segmentation of the surveillance video. Furthermore, for each monitoring sub-video, frames are randomly extracted, and pixel texture recognition processing is performed on several randomly selected frames to obtain the pixel texture feature information corresponding to each monitoring sub-video. This pixel texture feature information can be, but is not limited to, the pixel texture feature information of the background environment corresponding to the monitoring sub-video, thereby representing the pixel texture of the background environment of each monitoring sub-video. Then, based on the similarity of the pixel texture feature information between all monitoring sub-videos, which is the similarity between the pixel texture feature information of any two monitoring sub-videos, all monitoring sub-videos are divided into several monitoring sub-video sets. This ensures that the similarity of the pixel texture feature information between any two monitoring sub-videos in each monitoring sub-video set is within a preset similarity range, thereby guaranteeing that all monitoring sub-videos in each monitoring sub-video set have the same or similar screen content.

[0070] Preferably, in step S3, based on the second-screen feature information of each of the monitoring sub-video sets, each monitoring sub-video set is input into a matching neural network model for recognition to obtain the corresponding video screen recognition result; based on the video screen recognition result, all monitoring sub-videos are filtered to retain the corresponding valid monitoring sub-videos, including:

[0071] For each monitoring sub-video set, pixel contour recognition processing is performed on all monitoring sub-videos to obtain the pixel contour distribution information of each monitoring sub-video, which is used as the second image feature information. The pixel contour distribution information is compared with the image type information of each neural network model in the neural network model library, so that each monitoring sub-video set is input into the matching neural network model for recognition, thereby obtaining the corresponding video image recognition result.

[0072] The video image recognition results are analyzed for object recognition credibility to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility is then compared with a threshold to filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos.

[0073] In the above technical solution, pixel contour recognition processing is performed on all monitoring sub-videos under each monitoring sub-video set to obtain the pixel contour distribution information of each monitoring sub-video. This pixel contour distribution information is used to characterize the actions and shapes of all objects on the screen of the monitoring sub-video. Then, this pixel contour distribution information is compared with the image type information adapted to each of the neural network models in the neural network model library. Each monitoring sub-video set is then input into the matching neural network model for recognition, thereby obtaining the corresponding video image recognition result. This allows for finding the most suitable neural network model for each monitoring sub-video set based on the actions and shapes of all objects on the screen, ensuring that the found neural network model can accurately and comprehensively identify and analyze all monitoring sub-videos under the corresponding monitoring sub-video set. Furthermore, an object recognition credibility analysis is performed on the video image recognition results to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility can be, but is not limited to, the credibility value of the object's action and shape recognition results. Then, a threshold comparison is performed on the recognition credibility, and the monitoring sub-videos with recognition credibility greater than the preset credibility threshold are retained as valid monitoring sub-videos, thereby improving the credibility of the monitoring sub-videos and ensuring the accuracy of scene security recognition.

[0074] Preferably, in step S4, based on the video image recognition results of each of the valid monitoring sub-videos, the abnormal event occurrence attribute information of the target site is determined; based on the abnormal event occurrence attribute information, the monitoring video is processed for identification, including:

[0075] The video image recognition results of each of the valid monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding object inside the target site; wherein, the action behavior feature information includes the type of action issued by the object, the range of influence of the action, and the time of the action;

[0076] Based on the type of action and the scope of its impact issued by the object, determine whether an abnormal event has occurred at the target site; if so, use the corresponding action occurrence time information as the abnormal event occurrence attribute information.

[0077] Based on the action occurrence time information corresponding to the occurrence attribute information of the abnormal event, the timeline of the monitoring video is marked with time points, thereby marking and processing the occurrence time points of all abnormal events in the monitoring video.

[0078] In the above technical solution, the video image recognition results of each of the effective monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding objects within the target site. This allows for the spatial and temporal representation of the actions performed by all objects within the target site. Based on the action type and impact range information of the object, it is determined whether an abnormal event has occurred at the target site. For example, it is determined whether the action type belongs to a preset dangerous action type or whether the impact range exceeds a preset area. If so, an abnormal event is confirmed, and the corresponding action occurrence time information is used as the abnormal event occurrence attribute information to record the occurrence time of the abnormal event. Furthermore, based on the action occurrence time information corresponding to the abnormal event occurrence attribute information, time point marking processing is performed on the timeline of the monitoring video. This process marks the occurrence time points of all abnormal events in the monitoring video. Thus, when the monitoring video reaches the corresponding abnormal event occurrence time point during playback, the image is marked, providing a visual alert of the abnormal event.

[0079] See Figure 2 This is a schematic diagram of the framework of the adaptive scene matching security surveillance video artificial intelligence recognition system provided in an embodiment of the present invention. The adaptive scene matching security surveillance video artificial intelligence recognition system includes:

[0080] The target scene recognition module is used to detect the target scene and obtain the motion state data of objects within the target scene; based on the object motion state data, the scene attribute information of the target scene is obtained.

[0081] The video surveillance operation adjustment module is used to adjust the video surveillance operation status of the target site based on the scene attribute information.

[0082] The surveillance video segmentation module is used to divide the surveillance video into several surveillance sub-videos based on the feature information of the video surveillance operation status.

[0083] The surveillance video integration module is used to perform frame sampling analysis on all surveillance sub-videos to determine the first frame feature information of all surveillance sub-videos, thereby dividing all surveillance sub-videos into several surveillance sub-video sets.

[0084] The artificial intelligence recognition module is used to input each monitoring sub-video set into a matching neural network model for recognition based on the second-screen feature information of each of the monitoring sub-video sets, and obtain the corresponding video screen recognition results.

[0085] The surveillance video filtering module is used to filter all surveillance sub-videos based on the video image recognition results and retain the corresponding valid surveillance sub-videos.

[0086] The abnormal event attribute determination module is used to determine the abnormal event occurrence attribute information of the target site based on the video image recognition results of each of the valid monitoring sub-videos.

[0087] The surveillance video identification module is used to identify and process the surveillance video based on the attribute information of the abnormal event.

[0088] This adaptive scene-matching security monitoring video AI recognition system detects the motion state data of objects within the target site, determines scene attribute information, and adjusts the video monitoring operation state of the target site accordingly, improving the matching accuracy of the shooting. Based on the feature information of the video monitoring operation state and the first-frame features of the monitoring video, the system divides and classifies the monitoring video into several monitoring sub-video sets. Each monitoring sub-video set is input into a matching neural network model for recognition, resulting in video image recognition results. This allows different monitoring sub-videos to be matched and recognized by AI, thereby filtering and retaining the effective monitoring sub-videos. Based on the video image recognition results of all effective monitoring sub-videos, the system determines the abnormal event occurrence attribute information of the target site, and then processes and labels the monitoring video to achieve scene-based matching processing of the target site. This improves the flexibility and accuracy of security identification and analysis of monitoring videos, ensuring the overall efficiency and reliability of the monitoring video.

[0089] Preferably, the target scene recognition module is used to detect the target scene and obtain object motion state data within the target scene; based on the object motion state data, it obtains scene attribute information of the target scene, including:

[0090] Non-contact scanning detection is performed on the target site to obtain the motion trajectory data and motion speed data of all objects inside the target site, which are used as the motion state data of the objects.

[0091] By analyzing the motion trajectory data and the motion speed data, the probability information of the occurrence of objects and the object distribution density information of each sub-region under the target site are determined, and these are used as the scene attribute information.

[0092] This video surveillance operation adjustment module is used to adjust the video surveillance operation status of the target site based on the scene attribute information, including:

[0093] Based on the probability information of the object's appearance and the distribution density information of the object, adjust the shooting operation parameters corresponding to video surveillance of each sub-area under the target site; wherein, the shooting operation parameters include at least one of the shooting field of view and the shooting focal length.

[0094] In the aforementioned technical solutions, the flow of people or movement of vehicles at target sites such as public places are complex and variable. The movement and distribution of objects (people and / or objects) in different areas of the target site vary at different times. If a uniform video shooting method is always used when monitoring the target site, it is impossible to guarantee detailed global monitoring of different areas within the target site, and it is impossible to accurately identify the dynamic changes of objects within each area. Therefore, non-contact scanning detection, such as radar scanning or structured light scanning, is first performed on the target site to detect the movement trajectory and speed of all objects within the target site during the corresponding time period. This obtains the movement trajectory data and speed data of each object, thereby quantitatively determining the movement of objects in different areas of the target site. Then, statistical analysis is performed on the motion trajectory data and motion speed data to obtain the object occurrence probability information and object distribution density information for each sub-region under the target site. These sub-regions can be, but are not limited to, pre-divided grid sub-regions within the target site. The statistical calculation of the object occurrence probability and object distribution density (i.e., the number of objects per unit area) is a conventional technique in this field and will not be detailed here. Based on this object occurrence probability and distribution density information, the shooting field of view and / or shooting focal length for video surveillance of each sub-region under the target site are adjusted. For example, if the object occurrence probability within a sub-region exceeds a preset probability threshold, the shooting focal length for that sub-region is increased to achieve accurate shooting of objects within that sub-region; if the object distribution density within a sub-region exceeds a preset density threshold, the shooting field of view for that sub-region is increased so that all objects appearing within that sub-region can appear within the corresponding shooting field of view. This allows for global quantitative identification of the object presence within the target site, enabling targeted identification of different objects for subsequent assessment of the safety status within the scene.

[0095] Preferably, the surveillance video segmentation module is used to divide the surveillance video into several surveillance sub-videos based on the feature information of the video surveillance operation status, including:

[0096] Based on the time characteristics of the shooting operation parameters change during the adjustment of the video surveillance operation status, the entire video surveillance process of the target site is divided into several shooting time intervals; based on the start and end points of all shooting time intervals, the surveillance video is divided into several surveillance sub-videos.

[0097] This surveillance video integration module is used to perform frame sampling analysis on all surveillance sub-videos separately, determine the first-frame feature information of each surveillance sub-video, and thereby divide all surveillance sub-videos into several surveillance sub-video sets, including:

[0098] For each monitoring sub-video, frames are randomly extracted, and pixel texture recognition processing is performed on several randomly extracted frames to obtain the pixel texture feature information corresponding to each monitoring sub-video, which is used as the first image feature information. Based on the similarity of the pixel texture feature information between all monitoring sub-videos, all monitoring sub-videos are divided into several monitoring sub-video sets. Among them, the similarity of the pixel texture feature information between all monitoring sub-videos under each monitoring sub-video set satisfies the preset similarity distribution condition.

[0099] In the above technical solution, during the video surveillance shooting of the target site, the corresponding shooting field of view and / or shooting focal length are adjusted so that the shooting field of view and / or shooting focal length change over time. Based on the characteristic information of the shooting field of view and / or shooting focal length changing over time, the entire video surveillance process of the target site is divided into several shooting time intervals. In each shooting time interval, the fluctuation rate of the shooting field of view and shooting focal length is within the corresponding fluctuation rate range, which ensures that the image of the surveillance video corresponding to each shooting time interval is relatively stable. Furthermore, based on the start and end points of all shooting time intervals, the surveillance video is divided into several surveillance sub-videos, so that each surveillance sub-video corresponds to a shooting time interval, achieving effective division of the surveillance video. Additionally, frames are randomly extracted from each surveillance sub-video, and pixel texture recognition processing is performed on several randomly extracted frames to obtain the pixel texture feature information corresponding to each surveillance sub-video. This pixel texture feature information can be, but is not limited to, the pixel texture feature information of the background environment corresponding to the surveillance sub-video, thereby representing the pixel texture of the background environment of each surveillance sub-video. Based on the similarity of pixel texture features between all the monitoring sub-videos, which is the similarity between pixel texture features between any two monitoring sub-videos, all the monitoring sub-videos are divided into several monitoring sub-video sets. This ensures that the similarity of pixel texture features between any two monitoring sub-videos in each monitoring sub-video set is within a preset similarity range, thereby guaranteeing that all monitoring sub-videos in each monitoring sub-video set have the same or similar video content.

[0100] Preferably, the artificial intelligence recognition module, based on the second-screen feature information of each of the monitoring sub-video sets, inputs each monitoring sub-video set into a matching neural network model for recognition, and obtains the corresponding video screen recognition results, including:

[0101] For each monitoring sub-video set, pixel contour recognition processing is performed on all monitoring sub-videos to obtain the pixel contour distribution information of each monitoring sub-video, which is used as the second image feature information. The pixel contour distribution information is compared with the image type information of each neural network model in the neural network model library, so that each monitoring sub-video set is input into the matching neural network model for recognition, thereby obtaining the corresponding video image recognition result.

[0102] This surveillance video filtering module is used to filter all surveillance sub-videos based on the video image recognition results, retaining the corresponding valid surveillance sub-videos, including:

[0103] The video image recognition results are analyzed for object recognition credibility to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility is then compared with a threshold to filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos.

[0104] In the described technical solution, pixel contour recognition processing is performed on all monitoring sub-videos under a given monitoring sub-video set to obtain pixel contour distribution information for each monitoring sub-video. This pixel contour distribution information is used to characterize the actions and shapes of all objects in the monitoring sub-video's frame. This pixel contour distribution information is then compared with the image type information adapted to each neural network model in the neural network model library. Each monitoring sub-video set is then input into the matching neural network model for recognition, thereby obtaining the corresponding video frame recognition result. This approach allows for finding the most suitable neural network model for each monitoring sub-video set based on the actions and shapes of all objects in the frame, ensuring that the chosen neural network model can accurately and comprehensively identify and analyze all monitoring sub-videos within the corresponding monitoring sub-video set. Furthermore, an object recognition credibility analysis is performed on the video image recognition results to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility can be, but is not limited to, the credibility value of the object's action and shape recognition results. Then, a threshold comparison is performed on the recognition credibility, and the monitoring sub-videos with recognition credibility greater than the preset credibility threshold are retained as valid monitoring sub-videos, thereby improving the credibility of the monitoring sub-videos and ensuring the accuracy of scene security recognition.

[0105] Preferably, the abnormal event attribute determination module determines the abnormal event occurrence attribute information of the target site based on the video frame recognition results of each of the valid monitoring sub-videos, including:

[0106] The video image recognition results of each of the valid monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding object inside the target site; wherein, the action behavior feature information includes the type of action issued by the object, the range of influence of the action, and the time of the action;

[0107] Based on the type of action and the scope of its impact issued by the object, determine whether an abnormal event has occurred at the target site; if so, use the corresponding action occurrence time information as the abnormal event occurrence attribute information.

[0108] The surveillance video identification module performs identification processing on the surveillance video based on the attribute information of the abnormal event, including:

[0109] Based on the action occurrence time information corresponding to the occurrence attribute information of the abnormal event, the timeline of the monitoring video is marked with time points, thereby marking and processing the occurrence time points of all abnormal events in the monitoring video.

[0110] In the described technical solution, the video image recognition results of all valid monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding objects within the target site. This allows for the spatial and temporal representation of the actions performed by all objects within the target site. Based on the action type and impact range information of the object, it is determined whether an abnormal event has occurred at the target site. For example, it is determined whether the action type belongs to a preset dangerous action type or whether the impact range exceeds a preset area. If so, an abnormal event is confirmed, and the corresponding action occurrence time information is used as the abnormal event occurrence attribute information to record the occurrence time of the abnormal event. Furthermore, based on the action occurrence time information corresponding to the abnormal event occurrence attribute information, time point marking processing is performed on the timeline of the monitoring video. This process marks the occurrence time points of all abnormal events in the monitoring video. Thus, when the monitoring video reaches the corresponding abnormal event occurrence time point during playback, the image is marked, providing a visual alert of the abnormal event.

[0111] As can be seen from the above embodiments, the adaptive scene matching security monitoring video AI recognition method and system detects the motion state data of objects inside the target site, determines scene attribute information, and adjusts the video monitoring operation state of the target site accordingly, thereby improving the shooting matching of the target site in a targeted manner; based on the feature information of the video monitoring operation state and the first screen features of the monitoring video, the monitoring video is divided and classified into several monitoring sub-video sets, and each monitoring sub-video set is input into a matching neural network model for recognition, so as to obtain video screen recognition results, enabling different monitoring sub-videos to be matched by AI recognition, thereby filtering and retaining the effective monitoring sub-videos; based on the video screen recognition results of all effective monitoring sub-videos, the abnormal event occurrence attribute information of the target site is determined, and the monitoring video is then labeled and processed to achieve scene-based matching processing of the target site, improving the flexibility and accuracy of security recognition analysis of monitoring video, and ensuring the efficiency and reliability of monitoring video.

[0112] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-powered method for adaptive scene matching in security surveillance video recognition, characterized in that: It includes the following steps: Step S1: Detect the target scene to obtain object motion state data within the target scene; based on the object motion state data, obtain scene attribute information of the target scene; Based on the scene attribute information, adjust the video surveillance operation status of the target site; Step S2: Based on the feature information of the video monitoring operation status, the monitoring video is divided into several monitoring sub-videos; frame sampling analysis is performed on all monitoring sub-videos to determine the first frame feature information of all monitoring sub-videos, thereby dividing all monitoring sub-videos into several monitoring sub-video sets. Step S3: Based on the second screen feature information of each of the monitoring sub-video sets, input each monitoring sub-video set into the matching neural network model for recognition to obtain the corresponding video screen recognition result; based on the video screen recognition result, filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos; Step S4: Based on the video image recognition results of each of the valid monitoring sub-videos, determine the abnormal event occurrence attribute information of the target site; Based on the abnormal event occurrence attribute information, the monitoring video is labeled and processed.

2. The AI-powered security surveillance video recognition method with adaptive scene matching as described in claim 1, characterized in that: In step S1, the target scene is detected to obtain object motion state data within the target scene; based on the object motion state data, scene attribute information of the target scene is obtained. Based on the scene attribute information, adjust the video surveillance operation status of the target site, including: Non-contact scanning detection is performed on the target site to obtain the motion trajectory data and motion speed data of all objects inside the target site, which are used as the motion state data of the objects. The motion trajectory data and motion speed data are analyzed to determine the object occurrence probability information and object distribution density information for each of the sub-regions under the target site, which are used as the scene attribute information. Based on the object occurrence probability information and the object distribution density information, the shooting operation parameters corresponding to video monitoring of each sub-region under the target site are adjusted. The shooting operation parameters include at least one of the shooting field of view and the shooting focal length.

3. The AI-powered security surveillance video recognition method with adaptive scene matching as described in claim 1, characterized in that: In step S2, based on the feature information of the video surveillance operation status, the surveillance video is divided into several surveillance sub-videos; frame sampling analysis is performed on each surveillance sub-video to determine the first frame feature information of each surveillance sub-video, thereby dividing all surveillance sub-videos into several surveillance sub-video sets, including: Based on the time characteristic information of the shooting operation parameters change during the adjustment of the video monitoring operation status, the entire video monitoring process of the target site is divided into several shooting time intervals; based on the start and end points of all shooting time intervals, the monitoring video is divided into several monitoring sub-videos. For each monitoring sub-video, frames are randomly extracted, and pixel texture recognition processing is performed on several randomly extracted frames to obtain the pixel texture feature information corresponding to each monitoring sub-video, which is used as the first image feature information. Based on the similarity of the pixel texture feature information between all monitoring sub-videos, all monitoring sub-videos are divided into several monitoring sub-video sets. Among them, the similarity of the pixel texture feature information between all monitoring sub-videos under each monitoring sub-video set satisfies the preset similarity distribution condition.

4. The AI-powered method for adaptive scene matching in security surveillance video recognition as described in claim 1, characterized in that: In step S3, based on the second screen feature information of each of the monitoring sub-video sets, each monitoring sub-video set is input into a matching neural network model for recognition, and the corresponding video screen recognition result is obtained. Based on the video image recognition results, all monitoring sub-videos are filtered, and the corresponding valid monitoring sub-videos are retained, including: For each monitoring sub-video set, pixel contour recognition processing is performed on all monitoring sub-videos to obtain the pixel contour distribution information of each monitoring sub-video, which is used as the second image feature information. The pixel contour distribution information is compared with the image type information of each neural network model in the neural network model library, so that each monitoring sub-video set is input into the matching neural network model for recognition, thereby obtaining the corresponding video image recognition result. The video image recognition results are analyzed for object recognition credibility to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility is then compared with a threshold to filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos.

5. The AI-powered security surveillance video recognition method with adaptive scene matching as described in claim 1, characterized in that: In step S4, based on the video image recognition results of each of the valid monitoring sub-videos, the abnormal event occurrence attribute information of the target site is determined; Based on the abnormal event occurrence attribute information, the monitoring video is marked and processed, including: The video image recognition results of each of the valid monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding objects inside the target site; wherein, the action behavior feature information includes the type of action issued by the object, the range of influence of the action, and the time of the action; Based on the type of action and the scope of influence of the action issued by the object, it is determined whether an abnormal event has occurred at the target site; if so, the corresponding action occurrence time information is used as the abnormal event occurrence attribute information. Based on the action occurrence time information corresponding to the abnormal event occurrence attribute information, the timeline of the monitoring video is marked with time points, thereby marking and processing the monitoring video for all abnormal event occurrence time points.

6. An AI-powered security monitoring video recognition system with adaptive scene matching, characterized in that, include: The target scene recognition module is used to detect the target scene and obtain the motion state data of objects inside the target scene; Based on the object's motion state data, the scene attribute information of the target site is obtained; The video surveillance operation adjustment module is used to adjust the video surveillance operation status of the target site based on the scene attribute information. The surveillance video segmentation module is used to divide the surveillance video into several surveillance sub-videos based on the feature information of the video surveillance operation status. The surveillance video integration module is used to perform frame sampling analysis on all surveillance sub-videos to determine the first frame feature information of all surveillance sub-videos, thereby dividing all surveillance sub-videos into several surveillance sub-video sets. The artificial intelligence recognition module is used to input each monitoring sub-video set into a matching neural network model for recognition based on the second-screen feature information of each of the monitoring sub-video sets, and obtain the corresponding video screen recognition results. The surveillance video filtering module is used to filter all surveillance sub-videos based on the video image recognition results and retain the corresponding valid surveillance sub-videos. The abnormal event attribute determination module is used to determine the abnormal event occurrence attribute information of the target site based on the video image recognition results of each of the valid monitoring sub-videos. The surveillance video identification module is used to identify and process the surveillance video based on the abnormal event occurrence attribute information.

7. The adaptive scene matching security surveillance video artificial intelligence recognition system as described in claim 6, characterized in that: The target scene recognition module is used to detect the target scene and obtain the motion state data of objects inside the target scene; Based on the object's motion state data, the scene attribute information of the target site is obtained, including: Non-contact scanning detection is performed on the target site to obtain the motion trajectory data and motion speed data of all objects inside the target site, which are used as the motion state data of the objects. The motion trajectory data and the motion speed data are analyzed to determine the object occurrence probability information and object distribution density information corresponding to each of the sub-regions under the target site, which are then used as the scene attribute information. The video surveillance operation adjustment module is used to adjust the video surveillance operation status of the target site based on the scene attribute information, including: Based on the object occurrence probability information and the object distribution density information, the shooting operation parameters corresponding to video monitoring of each sub-area under the target site are adjusted; wherein, the shooting operation parameters include at least one of the shooting field of view range and the shooting focal length.

8. The adaptive scene matching security monitoring video artificial intelligence recognition system as described in claim 6, characterized in that: The surveillance video segmentation module is used to divide the surveillance video into several surveillance sub-videos based on the feature information of the video surveillance operation status, including: Based on the time characteristic information of the shooting operation parameters change during the adjustment of the video monitoring operation status, the entire video monitoring process of the target site is divided into several shooting time intervals; based on the start and end points of all shooting time intervals, the monitoring video is divided into several monitoring sub-videos. The surveillance video integration module is used to perform frame sampling analysis on all surveillance sub-videos to determine the first-frame feature information of each surveillance sub-video, thereby dividing all surveillance sub-videos into several surveillance sub-video sets, including: For each monitoring sub-video, frames are randomly extracted, and pixel texture recognition processing is performed on several randomly extracted frames to obtain the pixel texture feature information corresponding to each monitoring sub-video, which is used as the first image feature information. Based on the similarity of the pixel texture feature information between all monitoring sub-videos, all monitoring sub-videos are divided into several monitoring sub-video sets. Among them, the similarity of the pixel texture feature information between all monitoring sub-videos under each monitoring sub-video set satisfies the preset similarity distribution condition.

9. The adaptive scene matching security monitoring video artificial intelligence recognition system as described in claim 6, characterized in that: The artificial intelligence recognition module is used to input each monitoring sub-video set into a matching neural network model for recognition based on the second-screen feature information of each of the monitoring sub-video sets, and obtain the corresponding video screen recognition results, including: For each monitoring sub-video set, pixel contour recognition processing is performed on all monitoring sub-videos to obtain the pixel contour distribution information of each monitoring sub-video, which is used as the second image feature information. The pixel contour distribution information is compared with the image type information of each neural network model in the neural network model library, so that each monitoring sub-video set is input into the matching neural network model for recognition, thereby obtaining the corresponding video image recognition result. The surveillance video filtering module is used to filter all surveillance sub-videos based on the video image recognition results, retaining the corresponding valid surveillance sub-videos, including: The video image recognition results are analyzed for object recognition credibility to obtain the recognition credibility of all objects in the video image recognition results. The recognition credibility is then compared with a threshold to filter all monitoring sub-videos and retain the corresponding valid monitoring sub-videos.

10. The adaptive scene matching security monitoring video artificial intelligence recognition system as described in claim 6, characterized in that: The abnormal event attribute determination module is used to determine the abnormal event occurrence attribute information of the target site based on the video frame recognition results of each of the valid monitoring sub-videos, including: The video image recognition results of each of the valid monitoring sub-videos are analyzed to obtain the action behavior feature information of the corresponding objects inside the target site; wherein, the action behavior feature information includes the type of action issued by the object, the range of influence of the action, and the time of the action; Based on the type of action and the scope of influence of the action issued by the object, it is determined whether an abnormal event has occurred at the target site; if so, the corresponding action occurrence time information is used as the abnormal event occurrence attribute information. The surveillance video identification module is used to identify and process the surveillance video based on the abnormal event occurrence attribute information, including: Based on the action occurrence time information corresponding to the abnormal event occurrence attribute information, the timeline of the monitoring video is marked with time points, thereby marking and processing the monitoring video for all abnormal event occurrence time points.

Images