A dangerous event positioning method and system based on AI visual recognition

Abstract

The application is suitable for the field of safety monitoring, and provides a dangerous event positioning method and system based on AI visual recognition. The method continuously collects monitoring video streams through a main camera and calculates motion violation degree, generates metadata containing abnormal area coordinates and uncertainty heat map after triggering an anomaly; a three-dimensional space model is established and multiple-view collaborative cameras are dispatched for anomaly verification; multi-view geometric intersection calculation is performed in combination with the main camera and the collaborative camera feature points, the preliminary three-dimensional coordinates of the dangerous event are solved, and the observation confidence is evaluated; finally, spatial probability diffusion simulation is performed by fusing multiple source information to generate a two-dimensional probability heat map. Through AI visual recognition and multi-source data fusion, the application realizes accurate positioning and uncertainty quantification of dangerous events, improves positioning reliability and emergency disposal efficiency, and is suitable for various safety monitoring scenarios.

Description

Technical Field

[0001] This invention belongs to the field of security monitoring, and in particular relates to a method and system for locating dangerous events based on AI visual recognition. Background Technology

[0002] With the increasing demands for security and control in public safety, industrial production, and urban governance, video surveillance systems have been widely deployed, becoming one of the core infrastructures for security assurance. In recent years, the rapid development of AI visual recognition technology has driven the transformation of surveillance systems from traditional passive recording and playback to proactive anomaly detection and event early warning. It can identify potential dangerous events based on image features and motion patterns, providing early response support for security and control. Currently, the application of AI visual recognition in dangerous event monitoring covers multiple scenarios, including abnormal personnel behavior, equipment failure, and sudden safety accidents. Its core requirements focus on quickly detecting anomalies and accurately locating event sites, buying time for emergency response.

[0003] Existing technologies for hazard location are mostly based on video analysis from a single camera, using methods such as optical flow and background modeling to detect motion anomalies, or using pre-trained models to identify specific hazard categories. Some multi-camera solutions only achieve multi-source video acquisition through simple field-of-view coverage, relying on single-view image coordinate transformation for the location of abnormal areas. They lack a unified spatial modeling and multi-view collaborative verification mechanism, and the final location results are mostly output in the form of a single coordinate point, without considering various uncertainties in the location process.

[0004] Existing technologies are insufficient for accurate and reliable location of dangerous events. Due to limitations of single-view observation, environmental interference, and location uncertainty, problems such as abnormal misjudgment and large location deviations are prone to occur. It is impossible to accurately pinpoint the true spatial location of dangerous events, and it is also difficult to quantify the credibility of the location results. This leads to low efficiency in subsequent emergency response and fails to meet the actual needs of safety prevention and control for accurate and efficient location. Summary of the Invention

[0005] The purpose of this invention is to provide a method for locating dangerous events based on AI visual recognition, aiming to solve the technical problems existing in the prior art as identified in the background section.

[0006] This invention is implemented as follows: a method for locating dangerous events based on AI visual recognition, the method comprising:

[0007] The main camera deployed in the target monitoring area continuously collects monitoring video streams, and performs real-time analysis on the monitoring video streams to calculate the motion violation degree of the current image frame;

[0008] When the motion violation exceeds a preset threshold, anomaly detection is triggered, generating anomaly metadata containing the coordinates of the current anomaly region and the uncertainty heatmap of the current frame.

[0009] Establish a three-dimensional spatial model covering the target monitoring area. Based on the current abnormal area coordinates, identify at least one collaborative camera that overlaps with the main camera's field of view and has a different viewing angle. Send a collaborative analysis command containing the current abnormal area coordinates to at least one collaborative camera.

[0010] The system receives collaborative analysis data from the collaborative camera after scanning the space corresponding to the coordinates of the current abnormal area according to the collaborative analysis instructions. The collaborative analysis data includes collaborative view feature points identified under the collaborative view. When the collaborative analysis data confirms that there is an anomaly in the current abnormal area, the anomaly is determined as a dangerous event target to be located.

[0011] By combining the main view feature points in the surveillance video stream, multi-view geometric intersection calculation is performed to calculate the preliminary spatial coordinates of the target in the three-dimensional space and evaluate the confidence weight of the observation results of the main camera and the cooperative camera.

[0012] By combining preliminary spatial coordinates, confidence weights of observation results, and the current frame uncertainty heatmap in the abnormal metadata, spatial probability diffusion simulation is performed on the geographic information base map to generate a two-dimensional probability heatmap that characterizes the possible locations of dangerous event targets within the target monitoring area.

[0013] As a further aspect of the present invention, the continuous acquisition of monitoring video streams and the real-time analysis of the monitoring video streams specifically include:

[0014] The main camera captures continuous images at a fixed frame rate to form the raw data sequence of the surveillance video stream;

[0015] Optical flow analysis is performed on the current image frame in the monitoring video stream to track the displacement vectors of feature points between frames, and statistical outliers of the acceleration and direction change rate of the displacement vectors are calculated to obtain the motion violation degree.

[0016] As a further aspect of the present invention, the generation of anomaly metadata including the coordinates of the current anomaly region and the uncertainty heatmap of the current frame specifically includes:

[0017] The motion violation degree is compared with the preset motion violation threshold. When the motion violation degree exceeds the motion violation threshold, the current moment of the monitoring video stream is determined as the abnormal trigger moment, and the current image frame is marked as a key frame.

[0018] Based on the distribution of outlier values ​​in the displacement vector in optical flow analysis, the continuous region with the largest displacement vector outlier value in the key frame is located as the key image region.

[0019] Based on the obtained key image regions, a bounding box algorithm is used to define the regions, and the pixel coordinates of the bounding boxes of the key image regions are extracted as the coordinates of the current abnormal region.

[0020] A hazard identification neural network model is established, and forward propagation is performed on key frames to obtain the category probability distribution of each pixel belonging to several predetermined hazard-related categories. The confidence variance of the category probability distribution of each pixel is calculated, and a grayscale image of the same size as the key frame is generated, where the pixel grayscale value represents the degree of uncertainty at that position, thus obtaining the uncertainty heatmap of the current frame.

[0021] The coordinates of the current abnormal region and the uncertainty heatmap of the current frame are encapsulated into a data packet to generate abnormal metadata.

[0022] As a further aspect of the present invention, determining at least one collaborative camera that overlaps with the main camera's field of view but has a different viewing angle specifically includes:

[0023] Load the internal parameter matrix and external pose matrix of the main camera and all candidate cooperative cameras. The internal parameter matrix includes the focal length and principal point coordinates, and the external pose matrix defines the position and orientation of the camera relative to the world coordinate system.

[0024] Based on the external pose matrices of all main cameras and all candidate collaborative cameras, a common reference frame containing the target monitoring area is defined as the world coordinate system of the three-dimensional spatial model.

[0025] Based on the external pose matrix of the main camera and the coordinates of the current abnormal region, calculate the three-dimensional spatial cone corresponding to the current abnormal region in the world coordinate system.

[0026] Traverse all candidate collaborative cameras, calculate the intersection of the field of view range of each candidate camera and the field of view of the three-dimensional spatial cone in the world coordinate system, and calculate the angle between the optical axis direction of the candidate collaborative camera and the optical axis direction of the main camera.

[0027] Candidate cameras with non-empty field-of-view intersections and optical axis angles greater than a preset angle threshold are selected and identified as collaborative cameras.

[0028] As a further aspect of the present invention, when the collaborative analysis data confirms the existence of an anomaly in the current abnormal area, the anomaly is determined as a dangerous event target to be located, specifically including:

[0029] Control all collaborative cameras that receive collaborative analysis commands, adjust the gimbal and zoom, and align the center of the field of view with the three-dimensional spatial cone corresponding to the coordinates of the current abnormal area;

[0030] The system continuously receives collaborative images captured by the collaborative camera within the adjusted field of view, identifies collaborative feature points in the collaborative images, records the collaborative pixel coordinates of each collaborative feature point, and encapsulates them into a collaborative viewpoint feature point data packet as collaborative analysis data.

[0031] Based on the collaborative pixel coordinates of the collaborative feature points, the pixel block information of the corresponding local area in the collaborative image is extracted. The pixel block information is then compared with the pre-stored normal scene image blocks for similarity calculation. If the similarity calculation result is lower than the preset normal similarity threshold, it is determined that the collaborative analysis data confirms that there is an anomaly in the current abnormal area, and the anomaly is officially marked as a dangerous event target to be located.

[0032] As a further aspect of the present invention, the calculation of the preliminary spatial coordinates of the hazardous event target in three-dimensional space specifically includes:

[0033] From the keyframes of the monitoring video stream at the moment of abnormal triggering, extract the main view feature points within the local image block corresponding to the coordinates of the current abnormal area, and record the main view pixel coordinates of each main view feature point;

[0034] Geometric correspondence matching is performed between the main view feature points and their pixel coordinates and the co-view feature points and their pixel coordinates to form a set of successfully matched feature point pairs.

[0035] Based on the world coordinate system of the 3D spatial model, the internal parameter matrix and external pose matrix of the main camera and the cooperative camera, for each successfully matched feature point pair, calculate the 3D spatial coordinates of each pair of feature points in the world coordinate system.

[0036] Cluster analysis is performed on all three-dimensional spatial point coordinates, and the geometric center of the three-dimensional spatial point coordinates of the highest density cluster is calculated as the preliminary spatial coordinates of the hazardous event target.

[0037] The matching success rate is calculated based on the proportion of the number of successfully matched feature point pairs to the number of feature points in the main view, and the view difference is calculated based on the angle between the line-of-sight vectors of the main camera and the cooperating camera to the initial spatial coordinates.

[0038] Based on the matching success rate and the difference in viewing angle, the confidence weights of the observation results of the main camera and the cooperative camera are calculated respectively.

[0039] As a further aspect of the present invention, the generation of a two-dimensional probability heatmap for characterizing the possible locations of hazardous event targets within the target monitoring area specifically includes:

[0040] The initial spatial coordinates of the hazardous event target are transformed from the world coordinate system of the three-dimensional spatial model to a coordinate system consistent with the geographic information base. Figure 1By establishing a two-dimensional geographic coordinate system, we can obtain two-dimensional projected coordinates.

[0041] The current frame uncertainty heatmap in the abnormal metadata is transformed by coordinate transformation and mapped onto the geographic information base map to obtain an uncertainty distribution map in the geographic coordinate system.

[0042] Centered on the two-dimensional projected coordinates, and combined with the confidence weight of the observation results, a two-dimensional Gaussian kernel function is constructed, and a pixel-by-pixel probability multiplication operation is performed with the uncertainty distribution map under the geographic coordinate system to obtain the probability value of the dangerous event target appearing at each pixel point, thus simulating the spatial probability diffusion process.

[0043] The uncertainty distribution map is normalized and the probability values ​​are mapped to color gradients to generate the final two-dimensional probability heatmap. The color intensity of each pixel in the two-dimensional probability heatmap represents the probability density of the occurrence of the dangerous event target at that pixel.

[0044] Another object of the present invention is to provide a hazardous event location system based on AI visual recognition, the system comprising:

[0045] The monitoring video stream acquisition module is used to continuously acquire monitoring video streams through the main camera deployed in the target monitoring area, and to perform real-time analysis on the monitoring video streams to calculate the motion violation degree of the current image frame;

[0046] The metadata generation module is used to trigger anomaly detection and generate anomaly metadata containing the coordinates of the current anomaly area and the uncertainty heatmap of the current frame when the motion violation exceeds a preset threshold.

[0047] The collaborative camera analysis trigger module is used to establish a three-dimensional spatial model covering the target monitoring area. Based on the current abnormal area coordinates, it identifies at least one collaborative camera that overlaps with the main camera's field of view and has a different viewing angle, and sends a collaborative analysis command containing the current abnormal area coordinates to at least one collaborative camera.

[0048] The anomaly confirmation module is used to receive collaborative analysis data transmitted back by the collaborative camera after scanning the space corresponding to the coordinates of the current anomaly area according to the collaborative analysis instructions. The collaborative analysis data includes collaborative view feature points identified under the collaborative view. When the collaborative analysis data confirms that there is an anomaly in the current anomaly area, the anomaly is determined as a dangerous event target to be located.

[0049] The preliminary coordinate calculation module is used to combine the main view feature points in the surveillance video stream to perform multi-view geometric intersection calculation, calculate the preliminary spatial coordinates of the target of the dangerous event in three-dimensional space, and evaluate the confidence weight of the observation results of the main camera and the cooperative camera.

[0050] The two-dimensional probability heatmap generation module is used to combine preliminary spatial coordinates, confidence weights of observation results, and current frame uncertainty heatmaps in anomaly metadata to perform spatial probability diffusion simulation on a geographic information base map, generating a two-dimensional probability heatmap to characterize the possible locations of hazardous event targets within the target monitoring area.

[0051] The beneficial effects of this invention are:

[0052] This invention effectively solves the core problems of insufficient positioning accuracy and low reliability in existing technologies by constructing an integrated positioning system that combines multi-view collaborative analysis and 3D spatial modeling. The solution leverages real-time motion violation analysis from the main camera to quickly detect anomalies, combines this with multi-view verification from collaborative cameras to eliminate false anomalies, and achieves accurate conversion of hazardous events from 2D image features to 3D spatial location through 3D spatial geometric intersection calculation and probability diffusion simulation. Simultaneously, it quantifies positioning uncertainty and presents the results with a probability heatmap. The overall solution not only improves the accuracy and robustness of hazardous event positioning and reduces the impact of environmental interference and perspective limitations, but also intuitively presents the possible distribution range of hazardous events, providing clear and reliable decision-making basis for emergency response. It is applicable to various complex monitoring scenarios and significantly improves the proactive response capability and handling efficiency of security and prevention. Attached Figure Description

[0053] Figure 1 A flowchart illustrating a hazardous event localization method based on AI visual recognition, provided as an embodiment of the present invention;

[0054] Figure 2 This is a flowchart illustrating the continuous acquisition of monitoring video streams and the real-time analysis of the monitoring video streams provided in an embodiment of the present invention.

[0055] Figure 3 A flowchart for generating anomaly metadata including the coordinates of the current anomaly region and the uncertainty heatmap of the current frame, provided for embodiments of the present invention;

[0056] Figure 4 A flowchart for determining at least one collaborative camera that overlaps with the field of view of the main camera and has a different viewing angle, as provided in an embodiment of the present invention;

[0057] Figure 5 A flowchart for determining an anomaly as a target of a dangerous event to be located, provided in an embodiment of the present invention;

[0058] Figure 6 This is a flowchart for calculating the preliminary spatial coordinates of a hazardous event target in three-dimensional space, provided by an embodiment of the present invention.

[0059] Figure 7A flowchart for generating a two-dimensional probability heatmap to characterize the possible locations of hazardous event targets within a target monitoring area, provided in an embodiment of the present invention;

[0060] Figure 8 This is a structural block diagram of a hazardous event location system based on AI visual recognition, provided as an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] Figure 1 A flowchart of a hazardous event localization method based on AI visual recognition provided in an embodiment of the present invention is shown below. Figure 8 As shown, the method includes:

[0063] S100: The main camera deployed in the target monitoring area continuously collects monitoring video streams and performs real-time analysis on the monitoring video streams to calculate the motion violation degree of the current image frame;

[0064] The main camera needs to continuously capture continuous images of the target monitoring area at a preset fixed frame rate to form a sequential raw video stream data sequence. This stable acquisition mode can ensure the continuity and consistency of information between frames.

[0065] In the real-time analysis phase of the video stream, the core technology uses optical flow analysis to correlate the current image frame with adjacent frames, accurately tracking the displacement vectors of various feature points in the image. These feature points encompass image elements corresponding to various moving targets that may exist in the monitored scene. Based on the tracked displacement vectors, the acceleration and rate of change of direction of each feature point during its motion are further calculated. Statistical methods are used to quantify the degree of deviation of the motion parameters from the motion patterns under normal conditions, ultimately obtaining the core quantitative indicator of motion violation. Through precise capture and quantitative analysis of the target motion state in the monitored scene, abnormal motion signals that significantly differ from normal motion patterns are filtered out.

[0066] S200: When the motion violation exceeds a preset threshold, anomaly detection is triggered, and anomaly metadata containing the coordinates of the current anomaly area and the uncertainty heatmap of the current frame is generated.

[0067] Motion violation degree is a core indicator for quantifying the degree of abnormality in motion patterns. When its signal exceeds a preset threshold, it means that the motion state in the current image frame has deviated from the statistical range of the normal scene. At this time, triggering anomaly perception can accurately pinpoint the initial time node where a dangerous event may occur.

[0068] By identifying the time when an anomaly is triggered and marking keyframes, the most valuable core images for analysis can be selected from the continuous video stream. This avoids the consumption of computing resources by massive amounts of non-anomaly data during subsequent processing, ensuring that the analysis focuses on effective information.

[0069] Based on the distribution of displacement vector outliers obtained from optical flow analysis, the continuous region with the largest displacement vector outlier is located as the key image region. By using the spatial concentration characteristics of motion anomalies, the core activity range of dangerous targets is locked. Then, the bounding box algorithm is used to define the region and extract pixel coordinates, transforming the abstract motion anomaly into specific and locatable spatial location information.

[0070] The hazard identification neural network model performs forward propagation on keyframes to obtain the category probability distribution of each pixel. Then, it generates an uncertainty heatmap by calculating the confidence variance, which quantifies the reliability of the model's judgment on whether each pixel belongs to the hazard category. This allows the localization process to fully consider the fuzzy areas of the model's judgment and avoid errors caused by absolute decision-making.

[0071] S300: Establish a three-dimensional spatial model covering the target monitoring area; based on the current abnormal area coordinates, identify at least one collaborative camera that overlaps with the main camera's field of view and has a different viewing angle; and send a collaborative analysis command containing the current abnormal area coordinates to at least one collaborative camera.

[0072] To establish a three-dimensional spatial model covering the target monitoring area, it is necessary to rely on the external pose matrices of the main camera and all candidate cooperative cameras, and establish a unified world coordinate system by defining a common reference frame, so that the observation data of cameras in different positions and orientations have spatial consistency.

[0073] Based on the external pose matrix of the main camera and the coordinates of the current abnormal area, a three-dimensional spatial cone is calculated, which accurately maps the abnormal area in the two-dimensional image plane to the three-dimensional physical space, clearly defining the actual spatial range of the abnormality and breaking the limitation that single-view images cannot reflect the real spatial location.

[0074] When determining the collaborative cameras, all candidate devices are traversed, and devices meeting the criteria are selected by calculating the intersection of the field of view and the angle between the optical axes. The intersection of the field of view ensures that the collaborative cameras can cover the abnormal area, while the difference in the optical axis angle ensures complementary observation perspectives, avoiding data redundancy or information loss due to similar perspectives. A collaborative analysis command containing the coordinates of the abnormal area is sent to the selected collaborative cameras, directly guiding them to focus on the target area. This allows for accurate data collection without full-area scanning, shortening response time and reducing computational resource consumption.

[0075] S400, receive collaborative analysis data from the collaborative camera after scanning the space corresponding to the coordinates of the current abnormal area according to the collaborative analysis instruction. The collaborative analysis data includes collaborative view feature points identified under the collaborative view. When the collaborative analysis data confirms that there is an anomaly in the current abnormal area, the anomaly is determined as a dangerous event target to be located.

[0076] After receiving the collaborative analysis command containing the coordinates of the abnormal area, the collaborative camera adjusts the pan-tilt head to change the shooting angle and adjusts the zoom to enhance the clarity of the target area, so that the center of the field of view is accurately aligned with the three-dimensional spatial cone corresponding to the abnormal area, ensuring that the acquired collaborative image can focus on the key area and capture clear and distinguishable details.

[0077] The system continuously receives collaborative images captured by the collaborative camera within the adjusted field of view. Through feature recognition algorithms, collaborative viewpoint feature points are extracted from the images. These feature points cover key visual information such as the contour and texture of the target in the abnormal area. The collaborative pixel coordinates of each feature point are recorded and encapsulated into collaborative analysis data to form a structured set of multi-view feature information.

[0078] Based on the pixel coordinates of collaborative feature points, pixel block information of the corresponding local area is extracted. The texture, grayscale, edge and other features of the pixel block are compared with the pre-built normal scene image block sample library. The pixel block features of the monitored area are stable under normal scene, but the occurrence of dangerous events will break this stable state, resulting in significant differences between the pixel block features and normal samples. When the similarity result is lower than the preset threshold, it can be confirmed that there is a real danger in the abnormal area, and then the abnormality is officially marked as a dangerous event target to be located.

[0079] The S500 combines the main view feature points in the surveillance video stream to perform multi-view geometric intersection calculations, calculates the preliminary spatial coordinates of the target in the three-dimensional space, and evaluates the confidence weights of the observation results of the main camera and the cooperative camera.

[0080] From the keyframes at the moment of abnormal triggering in the surveillance video stream, focus on the local image blocks corresponding to the coordinates of the abnormal area to extract the main view feature points. These feature points carry key visual information of dangerous targets, and recording the pixel coordinates of the main view can provide accurate two-dimensional reference for cross-view matching.

[0081] Geometric correspondence is matched between the feature points and pixel coordinates of the main viewpoint and the feature points and pixel coordinates of the co-viewpoint. By using spatial geometric rules, feature point pairs that can reflect the same spatial position are selected, thus eliminating the misalignment interference of feature points under different viewpoints.

[0082] Based on a unified 3D spatial model world coordinate system, and combining the internal parameter matrices and external pose matrices of the main camera and cooperative cameras, the system calculates the coordinates of each successfully matched feature point pair. Utilizing the geometric constraints of multi-view observations, the system minimizes the projection error using the least squares method to obtain the 3D spatial coordinates of each feature point in the world coordinate system, thus achieving the transformation from 2D features to 3D space. Cluster analysis is then performed on all calculated 3D spatial coordinates, aggregating effective feature points into density clusters and eliminating discrete noise points. The geometric center of the largest density cluster is used as the preliminary spatial coordinates of the hazardous event target, making the coordinate results more closely match the target's true spatial location.

[0083] The matching success rate is obtained by dividing the number of successfully matched feature point pairs by the total number of feature points in the main viewpoint. The viewpoint difference is obtained by dividing the line-of-sight vectors from the main camera and the cooperating camera to the initial spatial coordinates. These two metrics together constitute the core indicators for evaluating the reliability of the observation data. Based on these two indicators, the confidence weights of the observation results from the main camera and the cooperating camera are calculated to quantify the credibility of data from different cameras.

[0084] The S600, combining preliminary spatial coordinates, confidence weights of observation results, and current frame uncertainty heatmaps in anomalous metadata, performs spatial probability diffusion simulations on a geographic information base map to generate a two-dimensional probability heatmap characterizing the possible locations of hazardous event targets within the target monitoring area.

[0085] After receiving collaborative analysis instructions, the collaborative camera adjusts its shooting position and zooms to enhance target details through gimbal adjustment, so that the center of the field of view is precisely aligned with the three-dimensional spatial cone corresponding to the abnormal area. This ensures that the acquired collaborative images can focus on key areas and capture clear visual information sufficient to support subsequent analysis.

[0086] The system continuously receives collaborative images within the adjusted field of view and extracts collaborative viewpoint feature points through feature recognition algorithms. These feature points cover key visual information such as the contour, texture, and edges of the target. The collaborative pixel coordinates are recorded and encapsulated into structured collaborative analysis data.

[0087] Pixel block information of the corresponding local area is extracted based on the pixel coordinates of collaborative feature points. The visual features carried by the pixel blocks are compared with a pre-built normal scene image block sample library. The pixel block features of the monitored area under normal scene have stable properties, while the occurrence of dangerous events will break this stability, resulting in significant differences between the pixel block features and normal samples. When the similarity result is lower than the preset threshold, it can be confirmed that there is a real danger in the abnormal area, and then the abnormality can be clearly marked as a dangerous event target to be located.

[0088] like Figure 2As shown, the continuous acquisition of monitoring video streams and the real-time analysis of the monitoring video streams specifically include:

[0089] S110 captures continuous images at a fixed frame rate through the main camera, forming the raw data sequence of the surveillance video stream;

[0090] S120, perform optical flow analysis on the current image frame in the monitoring video stream, track the displacement vector of feature points between frames, calculate the statistical outliers of the acceleration and direction change rate of the displacement vector, and obtain the motion violation degree.

[0091] Motion violation is obtained by weighted fusion of acceleration outliers and direction change rate outliers of the displacement vector, and is characterized by statistical outliers after Z-score normalization.

[0092] ;

[0093] Parameter explanation:

[0094] Motion violation degree represents the degree of deviation between the motion pattern of feature points in the current image frame and the normal pattern; the larger the value, the more abnormal it is. These are weighting coefficients, with a range of values. The weights used to balance acceleration anomalies and directional change rate anomalies; Z-score is used to standardize outliers of acceleration, representing the degree to which the acceleration of a single feature point deviates from the normal distribution; The Z-score normalized outlier value of the direction change rate represents the degree to which the motion direction of a single feature point deviates abruptly from the normal distribution.

[0095] in:

[0096] ;

[0097] ;

[0098] Parameter explanation:

[0099] For the first The magnitude of the acceleration vector of a feature point is obtained by dividing the difference between the displacement vectors of two adjacent frames by the frame interval. For the first The displacement vector of each feature point in the current frame, obtained from optical flow analysis, represents the inter-frame pixel coordinate change. For the first The displacement vector of each feature point in the previous frame; The video frame interval is fixed at 1. For the first The rate of change of the motion direction of each feature point is represented by the angle between the displacement vectors of the current frame and the previous frame; The statistical mean of the acceleration of all feature points under normal conditions is obtained by using surveillance videos without dangerous events as the training set. This represents the statistical standard deviation of the acceleration of all feature points under normal conditions. This is the statistical mean of the rate of change of the motion direction of all feature points under normal conditions; This represents the statistical standard deviation of the rate of change of the motion direction of all feature points under normal conditions.

[0100] like Figure 3 As shown, the generation of anomaly metadata, which includes the coordinates of the current anomaly region and the uncertainty heatmap of the current frame, specifically includes:

[0101] S210, compare the motion violation degree with the preset motion violation threshold. When the motion violation degree exceeds the motion violation threshold, determine the current moment of the monitoring video stream as the abnormal trigger moment, and mark the current image frame as a key frame.

[0102] During the offline training phase, a large number of surveillance videos without dangerous events were used to statistically determine the distribution range of motion violation in normal scenarios, and a preset threshold was established. The target value is the 99.7 quantile of this distribution, ensuring that the false trigger rate for normal movement is below 0.3%. Dangerous events inevitably involve sudden changes in acceleration or direction, leading to a significant deviation in movement from the target value. .

[0103] An exception is triggered only when the motion violation exceeds the threshold (default 2 consecutive frames), avoiding false triggers caused by single-frame noise, while ensuring real-time perception of dangerous events.

[0104] S220, Based on the distribution of outlier values ​​of displacement vectors in optical flow analysis, locate the continuous region with the largest outlier value of displacement vector in the key frame as the key image region;

[0105] Dangerous events tend to occur in a spatially concentrated manner, and their abnormal movements will form a continuous area in the image, which is the core activity range of the dangerous target.

[0106] Displacement anomalies in single pixels or isolated small regions may be caused by non-hazardous factors such as image sensor noise or rain / dust occlusion. However, anomalies in continuous regions are an inevitable result of hazardous targets occupying a certain image space, effectively filtering out isolated noise. The region with the largest displacement vector anomaly value corresponds to the most violently moving part of the hazardous event and is the key object for subsequent feature extraction and multi-view verification, which can reduce the computational cost of invalid regions.

[0107] By using connected component analysis to locate continuous regions, we can ensure that the spatial morphology of key image regions is consistent with that of real dangerous targets, thus providing a reasonable spatial range for subsequent boundary box delineation.

[0108] S230, Based on the obtained key image region, a bounding box algorithm is used to define the region, and the pixel coordinates of the bounding box of the key image region are extracted as the coordinates of the current abnormal region.

[0109] S240, establish a hazard identification neural network model, perform forward propagation on the key frame to obtain the category probability distribution of each pixel belonging to several predetermined hazard-related categories, calculate the confidence variance of the category probability distribution of each pixel, generate a grayscale image of the same size as the key frame, where the pixel grayscale value represents the degree of uncertainty at that position, and obtain the uncertainty heatmap of the current frame.

[0110] Confidence variance is used to quantify the dispersion of the class probability distribution of a single pixel, reflecting the uncertainty of the model's judgment on whether that pixel belongs to a dangerous class.

[0111] ;

[0112] in:

[0113] ;

[0114] Parameter explanation:

[0115] This represents the confidence variance for a single pixel. A higher value indicates less certainty in the model's classification of that pixel. The range is: ; To determine the number of hazard-related categories; For the first The probability of a hazard category, range: The keyframes are obtained by forward propagation of the keyframes using a hazard recognition neural network model. The average class probability of a single pixel, ranging from: ,Right now The arithmetic mean of the probabilities of each category.

[0116] S250 encapsulates the coordinates of the current abnormal region and the uncertainty heatmap of the current frame into a data packet to generate abnormal metadata.

[0117] like Figure 4 As shown, determining at least one collaborative camera that overlaps with the main camera's field of view but has a different viewing angle specifically includes:

[0118] S310, Load the internal parameter matrix and external pose matrix of the main camera and all candidate cooperative cameras. The internal parameter matrix includes the focal length and principal point coordinates. The external pose matrix defines the position and orientation of the camera relative to the world coordinate system.

[0119] S320 defines a common reference frame containing the target monitoring area as the world coordinate system of the three-dimensional spatial model, based on the external pose matrices of all main cameras and all candidate cooperative cameras.

[0120] S330, calculate the three-dimensional spatial cone corresponding to the current abnormal region in the world coordinate system based on the external pose matrix of the main camera and the coordinates of the current abnormal region;

[0121] Based on the pinhole camera model, there is a one-to-one projection relationship between pixel coordinates and spatial points in the world coordinate system. All pixels in the anomaly region correspond to a set of rays in the world coordinate system. The cone formed by these rays represents the actual coverage area of ​​the anomaly region in three-dimensional space. Specifically:

[0122] 1. Read the internal parameter matrix of the main camera. and external pose matrix , The rotation matrix for the main camera. Let be the translation vector of the main camera, where the internal parameter matrix of the main camera is denoted as . Characterizes the camera's own optical properties. The focal length of the camera along the x-axis is obtained from camera calibration experiments; The focal length of the camera along the y-axis;

[0123] 2. Set the bounding box pixel coordinates of the current abnormal region. Convert to image coordinate system coordinates (Taking the principal point of the image as the origin), the conversion formula is: , These are the coordinates of the camera's principal point, i.e., the position of the origin of the image coordinate system in the pixel coordinate system;

[0124] 3. Based on the camera pinhole model, first, the image coordinate system coordinates... Convert to normalized coordinates in the camera coordinate system: , The normalized coordinates are then converted to ray direction vectors in the world coordinate system. ;

[0125] 4. Using the position of the main camera in the world coordinate system as the vertex and all rays as generatrices, a three-dimensional spatial cone is formed to cover the anomaly area. The vertex of the cone... (The position of the main camera in the world coordinate system), the bus is for all satisfying The directional rays, the cone boundary is formed by the rays corresponding to the four corner points of the anomaly region boundary box, This is the ray direction vector in the world coordinate system, representing the spatial ray direction corresponding to the pixel. The vertex of the three-dimensional spatial cone represents the installation position of the main camera in the world coordinate system.

[0126] S340, traverse all candidate collaborative cameras, calculate the intersection of the field of view range of each candidate camera and the field of view of the three-dimensional spatial cone in the world coordinate system, and calculate the angle between the optical axis direction of the candidate collaborative camera and the optical axis direction of the main camera.

[0127] For the intersection of the fields of view:

[0128] 1. Read the internal parameter matrix of all candidate collaborative cameras. External pose matrix and field of view (Horizontal field of view) Vertical field of view );

[0129] 2. Based on and Define the field of view (pyramidal structure) of the candidate camera in the world coordinate system, the boundary of which is determined by the optical axis direction and the field of view angle:

[0130] Horizontal boundary: optical axis direction ± ;

[0131] Vertical boundary: Optical axis direction ± ;

[0132] 3. Extract the four boundary generatrices of the three-dimensional spatial cone of the main camera (corresponding to the four corner points of the bounding box of the abnormal region);

[0133] 4. For each generatrix, determine whether it falls within the field of view of the candidate camera (by checking whether the angle between the generatrix direction vector and the optical axis of the candidate camera is ≤). determination);

[0134] 5. If at least one busbar falls within the field of view of the candidate camera, then the field of view intersection is determined to be non-empty.

[0135] Regarding the angle between the optical axes:

[0136] 1. Extract the optical axis direction vector:

[0137] Main camera optical axis direction vector (Camera coordinate system optical axis is) (axis, converted to world coordinate system);

[0138] Candidate camera optical axis direction vector ;

[0139] 2. Calculate the angle between the two optical axes based on the vector dot product formula:

[0140] ;

[0141] in , For a unit vector, the formula simplifies to .

[0142] S350 filters out all candidate cameras whose field of view intersection is not empty and whose optical axis angle is greater than a preset angle threshold, and identifies them as collaborative cameras.

[0143] The observation data of the same spatial point from cameras with different perspectives can be used to calculate the three-dimensional coordinates of the spatial point through geometric intersection. The greater the difference in perspective, the higher the accuracy of depth estimation. At the same time, it is necessary to ensure that the collaborative cameras can cover abnormal areas and avoid blind spots.

[0144] Screening criteria:

[0145] 1. The field of view of the collaborative camera must cover the abnormal area (three-dimensional spatial cone) detected by the main camera; otherwise, effective image data of the abnormal area cannot be obtained, resulting in the failure of multi-view verification and positioning.

[0146] 2. The optical axis angle reflects the difference in viewing angle. When the angle is less than 30°, the viewing angles of the two cameras are similar, the redundancy of the observation data is high, and the depth estimation error of triangulation will increase significantly (the error is inversely proportional to the sine of the angle). When the angle is greater than or equal to 30°, the difference in viewing angle is sufficient to provide complementary depth information to ensure the accuracy of 3D positioning.

[0147] like Figure 5 As shown, when the collaborative analysis data confirms an anomaly in the current abnormal area, the anomaly is identified as a dangerous event target to be located, specifically including:

[0148] S410 controls all collaborative cameras that have received collaborative analysis commands, adjusts the gimbal and zoom, and aligns the center of the field of view with the three-dimensional spatial cone corresponding to the coordinates of the current abnormal area.

[0149] S420 continuously receives collaborative images captured by the collaborative camera within the adjusted field of view, identifies collaborative feature points in the collaborative images, records the collaborative pixel coordinates of each collaborative feature point, and encapsulates them into a collaborative viewpoint feature point data packet as collaborative analysis data.

[0150] S430, based on the collaborative pixel coordinates of the collaborative feature points, extract the pixel block information of the corresponding local area in the collaborative image, and perform similarity calculation between the pixel block information and the pre-stored normal scene image blocks. If the similarity calculation result is lower than the preset normal similarity threshold, it is determined that the collaborative analysis data confirms that there is an anomaly in the current abnormal area, and the anomaly is officially marked as a dangerous event target to be located.

[0151] When there are no dangerous events, local image patches in the monitored area have fixed texture, grayscale, and edge features, and are highly similar to pre-stored normal image patches; dangerous events will disrupt this stability, resulting in a significant decrease in similarity.

[0152] In the offline phase, normal scene images of the target monitoring area are collected at different times and under different lighting conditions. A large number of normal image blocks are cropped to generate normal image blocks (the size is consistent with the pixel blocks extracted from the collaborative images, such as 32×32 pixels) to build a normal sample library.

[0153] Structural similarity (SSIM) is used as the metric. SSIM can simultaneously consider three major features: brightness, contrast, and structure, making it more suitable for scene-level image patch similarity determination. The formula is:

[0154] ;

[0155] in: For coordinated image patches, This is a normal sample image patch. The mean, For variance, For covariance, It is a constant.

[0156] Statistically analyze the similarity distribution of all image patches in the normal sample library, with a preset threshold. The quintile of this distribution (i.e., 95% of normal image patches have a similarity ≥ 5%) ), ensuring a false positive rate of less than 5% in normal scenes; when the maximum similarity between the collaborative image patch and all normal sample image patches is <= When this occurs, it is considered abnormal.

[0157] like Figure 6 As shown, the calculation of the preliminary spatial coordinates of the hazardous event target in three-dimensional space specifically includes:

[0158] S510, extract the main view feature points in the local image block corresponding to the current abnormal area coordinates from the key frame of the monitoring video stream at the time of abnormal triggering, and record the main view pixel coordinates of each main view feature point.

[0159] S520 performs geometric correspondence matching between the main view feature points and the main view pixel coordinates and the co-view feature points and the co-view pixel coordinates to form a set of successfully matched feature point pairs.

[0160] S530, based on the world coordinate system of the three-dimensional spatial model, the internal parameter matrix and external pose matrix of the main camera and the cooperative camera, calculates the three-dimensional spatial coordinates of each pair of feature points in the world coordinate system for each successfully matched pair of feature points.

[0161] By matching feature point pairs between the main camera and the collaborative camera, the coordinates of three-dimensional points in the world coordinate system are calculated. The projection error is minimized using the least squares method, as shown in the following formula:

[0162] ;

[0163] The coordinates of the point in three-dimensional space are obtained after solving. .

[0164] in: , representing the scale factor, indicating the distance from a 3D point to the camera. (3D point to main camera coordinate system) (distance between axes) (3D point to collaborative camera coordinate system) (distance between axes) To match the coordinates of feature points in the three-dimensional space of the world coordinate system, i.e. the coordinates of the local feature points of the target of the dangerous event; Internal parameter matrices for the main camera and cooperating cameras; A rotation matrix for the main camera and cooperating cameras; Translation vectors for the main camera and cooperating cameras; The pixel coordinate vector of the feature points from the main viewpoint. ; The pixel coordinate vector of the feature points from the collaborative viewpoint. .

[0165] S540 performs cluster analysis on all three-dimensional spatial point coordinates and calculates the geometric center of the three-dimensional spatial point coordinates of the highest density cluster as the preliminary spatial coordinates of the hazardous event target.

[0166] S550 calculates the matching success rate based on the proportion of the number of successfully matched feature point pairs to the number of feature points in the main view, and calculates the view difference based on the angle between the line-of-sight vectors of the main camera and the cooperative camera to the initial spatial coordinates.

[0167] The difference in perspective is the angle between the line-of-sight vectors from the main camera / cooperative camera to the initial spatial coordinates, reflecting the complementarity of the perspectives of the two cameras observing the 3D points:

[0168] ;

[0169] in:

[0170] ;

[0171] The larger the value, the stronger the complementarity of the observation perspectives of the two cameras; The line-of-sight vectors from the main camera and the cooperating cameras to the initial spatial coordinates; The initial spatial coordinates of the hazardous event targets, i.e., the coordinates of the geometric center after clustering; The positions of the main camera and the cooperating cameras in the world coordinate system. , ;

[0172] S560 calculates the confidence weights of the observation results of the main camera and the cooperative camera based on the matching success rate and the difference in viewing angle.

[0173] The confidence score weights take into account both feature matching success rate and perspective differences, and are calculated using a normalized weighting method.

[0174] ;

[0175] Parameter explanation:

[0176] The confidence weights of the observation results from the main camera and the cooperating cameras are determined by weights; the higher the weight, the more reliable the observation data from that camera. To improve the feature matching success rate, , The number of successfully matched feature point pairs, Total number of feature points from the main viewpoint; The adjustment coefficient is used to adjust the influence of matching success rate and perspective difference on the weights, and is obtained through offline experiments for optimization.

[0177] like Figure 7 As shown, the generation of a two-dimensional probability heatmap to characterize the possible locations of hazardous event targets within the target monitoring area specifically includes:

[0178] S610 transforms the initial spatial coordinates of the hazardous event target from the world coordinate system of the three-dimensional spatial model to a coordinate system consistent with the geographic information base. Figure 1 By establishing a two-dimensional geographic coordinate system, we can obtain two-dimensional projected coordinates.

[0179] S620, perform coordinate transformation on the current frame uncertainty heatmap in the abnormal metadata and map it onto the geographic information base map to obtain an uncertainty distribution map under the geographic coordinate system;

[0180] S630, with the two-dimensional projected coordinates as the center, combined with the confidence weight of the observation results, constructs a two-dimensional Gaussian kernel function, and performs a pixel-by-pixel probability multiplication operation with the uncertainty distribution map under the geographic coordinate system to obtain the probability value of the dangerous event target appearing at each pixel point, simulating the spatial probability diffusion process.

[0181] S640 normalizes the uncertainty distribution map and maps the probability values ​​to color gradients to generate the final two-dimensional probability heatmap. The color intensity of each pixel in the two-dimensional probability heatmap represents the probability density of the occurrence of the dangerous event target at that pixel.

[0182] Hazard location suffers from multi-source errors. Two-dimensional probability heatmaps intuitively represent the probability of a hazardous event at each location through probability density distribution, avoiding the absolute errors of traditional single-point location.

[0183] The color gradient of the heat map can quickly guide staff to focus on high-probability areas, improving the efficiency of emergency response; at the same time, it supports multi-target positioning, with multiple high-probability peaks corresponding to multiple dangerous targets.

[0184] Specific procedures for locating hazardous events:

[0185] 1. Initial three-dimensional spatial coordinates Convert to two-dimensional geographic coordinate system using projection matrix :

[0186] ;

[0187] This is the projection matrix from the three-dimensional world coordinate system to the two-dimensional geographic coordinate system, provided by the GIS system.

[0188] 2. Create an uncertainty heatmap (pixel level) for the current frame. By mapping the coordinates (pixel coordinates → geographic coordinates) to the GIS base map, an uncertainty distribution map in the geographic coordinate system is obtained. .

[0189] 3. with Centered on, combined with confidence weights Construct a two-dimensional Gaussian kernel function:

[0190] ;

[0191] in , The base diffusion radius is 5 meters by default. The larger the value, the smaller the diffusion radius, and the higher the certainty of location.

[0192] 4. Perform pixel-by-pixel multiplication between the Gaussian kernel function and the uncertainty distribution map to obtain an initial probability map, and then normalize the initial probability map (the sum of the probabilities of all pixels is 1):

[0193] ;

[0194] ( (This indicates that the lower the uncertainty, the higher the probability weight).

[0195] 5. The probability is mapped to a color gradient (the higher the probability, the darker the color), generating a two-dimensional probability heatmap. During localization, the geometric center of the connected region (area ≥ preset threshold) with the highest probability density in the heatmap is taken as the final localization coordinate of the dangerous event, and the probability value of that region is output.

[0196] Figure 8 A structural block diagram of a hazardous event localization system based on AI visual recognition provided in an embodiment of the present invention is shown below. Figure 8 As shown, the system includes:

[0197] The monitoring video stream acquisition module 100 is used to continuously acquire monitoring video streams through a main camera deployed in the target monitoring area, and to perform real-time analysis on the monitoring video streams to calculate the motion violation degree of the current image frame;

[0198] Metadata generation module 200 is used to trigger anomaly detection and generate anomaly metadata containing the coordinates of the current anomaly area and the uncertainty heatmap of the current frame when the motion violation exceeds a preset threshold.

[0199] The collaborative camera analysis trigger module 300 is used to establish a three-dimensional spatial model covering the target monitoring area, determine at least one collaborative camera that overlaps with the main camera's field of view and has a different viewing angle based on the current abnormal area coordinates, and send a collaborative analysis command containing the current abnormal area coordinates to at least one collaborative camera.

[0200] The anomaly confirmation module 400 is used to receive collaborative analysis data transmitted back by the collaborative camera after scanning the space corresponding to the coordinates of the current anomaly area according to the collaborative analysis instruction. The collaborative analysis data includes collaborative view feature points identified under the collaborative view. When the collaborative analysis data confirms that there is an anomaly in the current anomaly area, the anomaly is determined as a dangerous event target to be located.

[0201] The preliminary coordinate calculation module 500 is used to combine the main view feature points in the monitoring video stream to perform multi-view geometric intersection calculation, calculate the preliminary spatial coordinates of the target of the dangerous event in three-dimensional space, and evaluate the confidence weight of the observation results of the main camera and the cooperative camera.

[0202] The two-dimensional probability heatmap generation module 600 is used to combine preliminary spatial coordinates, confidence weights of observation results, and current frame uncertainty heatmaps in abnormal metadata to perform spatial probability diffusion simulation on a geographic information base map, generating a two-dimensional probability heatmap to characterize the possible locations of dangerous event targets within the target monitoring area.

[0203] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0204] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0205] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for locating hazardous events based on AI visual recognition, characterized in that, The method includes: The main camera deployed in the target monitoring area continuously collects monitoring video streams, and performs real-time analysis on the monitoring video streams to calculate the motion violation degree of the current image frame; When the motion violation exceeds a preset threshold, anomaly detection is triggered, generating anomaly metadata containing the coordinates of the current anomaly region and the uncertainty heatmap of the current frame. Establish a three-dimensional spatial model covering the target monitoring area. Based on the current abnormal area coordinates, identify at least one collaborative camera that overlaps with the main camera's field of view and has a different viewing angle. Send a collaborative analysis command containing the current abnormal area coordinates to at least one collaborative camera. The system receives collaborative analysis data from the collaborative camera after scanning the space corresponding to the coordinates of the current abnormal area according to the collaborative analysis instructions. The collaborative analysis data includes collaborative view feature points identified under the collaborative view. When the collaborative analysis data confirms that there is an anomaly in the current abnormal area, the anomaly is determined as a dangerous event target to be located. By combining the main view feature points in the surveillance video stream, multi-view geometric intersection calculation is performed to calculate the preliminary spatial coordinates of the target in the three-dimensional space and evaluate the confidence weight of the observation results of the main camera and the cooperative camera. By combining preliminary spatial coordinates, confidence weights of observation results, and the current frame uncertainty heatmap in the abnormal metadata, spatial probability diffusion simulation is performed on the geographic information base map to generate a two-dimensional probability heatmap that characterizes the possible locations of dangerous event targets within the target monitoring area.

2. The method according to claim 1, characterized in that, The continuous acquisition and real-time analysis of the monitoring video stream specifically includes: The main camera captures continuous images at a fixed frame rate to form the raw data sequence of the surveillance video stream; Optical flow analysis is performed on the current image frame in the monitoring video stream to track the displacement vectors of feature points between frames, and statistical outliers of the acceleration and direction change rate of the displacement vectors are calculated to obtain the motion violation degree.

3. The method according to claim 2, characterized in that, The generation of anomaly metadata, which includes the coordinates of the current anomaly region and the uncertainty heatmap of the current frame, specifically includes: The motion violation degree is compared with the preset motion violation threshold. When the motion violation degree exceeds the motion violation threshold, the current moment of the monitoring video stream is determined as the abnormal trigger moment, and the current image frame is marked as a key frame. Based on the distribution of outlier values ​​in the displacement vector in optical flow analysis, the continuous region with the largest displacement vector outlier value in the key frame is located as the key image region. Based on the obtained key image regions, a bounding box algorithm is used to define the regions, and the pixel coordinates of the bounding boxes of the key image regions are extracted as the coordinates of the current abnormal region. A hazard identification neural network model is established, and forward propagation is performed on key frames to obtain the category probability distribution of each pixel belonging to several predetermined hazard-related categories. The confidence variance of the category probability distribution of each pixel is calculated, and a grayscale image of the same size as the key frame is generated, where the pixel grayscale value represents the degree of uncertainty at that position, thus obtaining the uncertainty heatmap of the current frame. The coordinates of the current abnormal region and the uncertainty heatmap of the current frame are encapsulated into a data packet to generate abnormal metadata.

4. The method according to claim 3, characterized in that, The determination of at least one collaborative camera that overlaps with the main camera's field of view but has a different viewing angle specifically includes: Load the internal parameter matrix and external pose matrix of the main camera and all candidate cooperative cameras. The internal parameter matrix includes the focal length and principal point coordinates, and the external pose matrix defines the position and orientation of the camera relative to the world coordinate system. Based on the external pose matrices of all main cameras and all candidate collaborative cameras, a common reference frame containing the target monitoring area is defined as the world coordinate system of the three-dimensional spatial model. Based on the external pose matrix of the main camera and the coordinates of the current abnormal region, calculate the three-dimensional spatial cone corresponding to the current abnormal region in the world coordinate system. Traverse all candidate collaborative cameras, calculate the intersection of the field of view of each candidate camera and the field of view of the three-dimensional spatial cone in the world coordinate system, and calculate the angle between the optical axis direction of the candidate collaborative camera and the optical axis direction of the main camera. Candidate cameras with non-empty field-of-view intersections and optical axis angles greater than a preset angle threshold are selected and identified as collaborative cameras.

5. The method according to claim 4, characterized in that, When the collaborative analysis data confirms an anomaly in the current abnormal area, the anomaly is identified as a hazardous event target to be located, specifically including: Control all collaborative cameras that receive collaborative analysis commands, adjust the gimbal and zoom, and align the center of the field of view with the three-dimensional spatial cone corresponding to the coordinates of the current abnormal area; The system continuously receives collaborative images captured by the collaborative camera within the adjusted field of view, identifies collaborative feature points in the collaborative images, records the collaborative pixel coordinates of each collaborative feature point, and encapsulates them into a collaborative viewpoint feature point data packet as collaborative analysis data. Based on the collaborative pixel coordinates of the collaborative feature points, the pixel block information of the corresponding local area in the collaborative image is extracted. The pixel block information is then compared with the pre-stored normal scene image blocks for similarity calculation. If the similarity calculation result is lower than the preset normal similarity threshold, it is determined that the collaborative analysis data confirms that there is an anomaly in the current abnormal area, and the anomaly is officially marked as a dangerous event target to be located.

6. The method according to claim 5, characterized in that, The calculation of the preliminary spatial coordinates of the hazardous event target in three-dimensional space specifically includes: From the keyframes of the monitoring video stream at the moment of abnormal triggering, extract the main view feature points within the local image block corresponding to the coordinates of the current abnormal area, and record the main view pixel coordinates of each main view feature point; Geometric correspondence matching is performed between the main view feature points and their pixel coordinates and the co-view feature points and their pixel coordinates to form a set of successfully matched feature point pairs. Based on the world coordinate system of the 3D spatial model, the internal parameter matrix and external pose matrix of the main camera and the cooperative camera, for each successfully matched feature point pair, calculate the 3D spatial coordinates of each pair of feature points in the world coordinate system. Cluster analysis is performed on all three-dimensional spatial point coordinates, and the geometric center of the three-dimensional spatial point coordinates of the highest density cluster is calculated as the preliminary spatial coordinates of the hazardous event target. The matching success rate is calculated based on the proportion of the number of successfully matched feature point pairs to the number of feature points in the main view, and the view difference is calculated based on the angle between the line-of-sight vectors of the main camera and the cooperating camera to the initial spatial coordinates. Based on the matching success rate and the difference in viewing angle, the confidence weights of the observation results of the main camera and the cooperative camera are calculated respectively.

7. The method according to claim 6, characterized in that, The generation of the two-dimensional probability heatmap, used to characterize the possible locations of hazardous event targets within the target monitoring area, specifically includes: The initial spatial coordinates of the hazardous event target are transformed from the world coordinate system of the three-dimensional spatial model to a two-dimensional geographic coordinate system consistent with the geographic information base map, thus obtaining two-dimensional projected coordinates; The current frame uncertainty heatmap in the abnormal metadata is transformed by coordinate transformation and mapped onto the geographic information base map to obtain an uncertainty distribution map in the geographic coordinate system. Centered on the two-dimensional projected coordinates, and combined with the confidence weight of the observation results, a two-dimensional Gaussian kernel function is constructed, and a pixel-by-pixel probability multiplication operation is performed with the uncertainty distribution map under the geographic coordinate system to obtain the probability value of the dangerous event target appearing at each pixel point, thus simulating the spatial probability diffusion process. The uncertainty distribution map is normalized and the probability values ​​are mapped to color gradients to generate the final two-dimensional probability heatmap. The color intensity of each pixel in the two-dimensional probability heatmap represents the probability density of the occurrence of the dangerous event target at that pixel.

8. A hazardous event location system based on AI visual recognition, characterized in that, The system includes: The monitoring video stream acquisition module is used to continuously acquire monitoring video streams through the main camera deployed in the target monitoring area, and to perform real-time analysis on the monitoring video streams to calculate the motion violation degree of the current image frame; The metadata generation module is used to trigger anomaly detection and generate anomaly metadata containing the coordinates of the current anomaly area and the uncertainty heatmap of the current frame when the motion violation exceeds a preset threshold. The collaborative camera analysis trigger module is used to establish a three-dimensional spatial model covering the target monitoring area. Based on the current abnormal area coordinates, it identifies at least one collaborative camera that overlaps with the main camera's field of view and has a different viewing angle, and sends a collaborative analysis command containing the current abnormal area coordinates to at least one collaborative camera. The anomaly confirmation module is used to receive collaborative analysis data transmitted back by the collaborative camera after scanning the space corresponding to the coordinates of the current anomaly area according to the collaborative analysis instructions. The collaborative analysis data includes collaborative view feature points identified under the collaborative view. When the collaborative analysis data confirms that there is an anomaly in the current anomaly area, the anomaly is determined as a dangerous event target to be located. The preliminary coordinate calculation module is used to combine the main view feature points in the surveillance video stream to perform multi-view geometric intersection calculation, calculate the preliminary spatial coordinates of the target of the dangerous event in three-dimensional space, and evaluate the confidence weight of the observation results of the main camera and the cooperative camera. The two-dimensional probability heatmap generation module is used to combine preliminary spatial coordinates, confidence weights of observation results, and current frame uncertainty heatmaps in anomaly metadata to perform spatial probability diffusion simulation on a geographic information base map, generating a two-dimensional probability heatmap to characterize the possible locations of hazardous event targets within the target monitoring area.

Images