A multi-camera linkage perimeter intrusion intelligent recognition and tracking method and system

Abstract

The application provides a multi-camera linkage perimeter intrusion intelligent recognition and tracking method and system, and relates to the technical field of video monitoring and intelligent security. In the application, the moving target in the multi-path picture is converted and fused into the actual position in the target area according to the camera installation pose calibration relationship; the intrusion is confirmed by crossing the forbidden area boundary, the number of continuous falling and the inward span; the takeover camera is selected according to the field of view connection relationship, the predicted arrival time and the preset position rotation time consumption, and the tracking number is maintained by using the target appearance characteristics; the cross-camera track is formed by splicing the single-camera tracking section and the cross-camera connection section, and the alarm instruction is output and updated according to the target position and the distance between the key protection positions; the technical scheme provided by the application not only effectively suppresses the instantaneous false alarm through the unified multi-camera space coordinates and multi-condition intrusion confirmation, but also improves the identity continuity of the cross-camera tracking and the traceability of the alarm track through the dynamic selection of the takeover opportunity and the appearance characteristic maintenance.

Description

Technical Field

[0001] This application relates to the fields of video surveillance and intelligent security technology, and in particular to a method and system for intelligent identification and tracking of perimeter intrusion using multi-camera linkage. Background Technology

[0002] In continuous monitoring scenarios such as parks, stations, transportation facilities, and other areas with perimeter restricted zones, multiple fixed and rotating cameras are typically deployed to monitor the perimeter. When a moving target enters the restricted area, the system needs to promptly identify the intrusion and continuously track the target's trajectory so that security personnel can accurately grasp the full picture of the intrusion event and take appropriate measures.

[0003] In existing technologies, most perimeter intrusion detection solutions perform boundary crossing detection or area intrusion determination within a single video feed. When the target leaves the current camera's field of view, it is necessary to manually switch the feed or have the system switch to an adjacent camera to continue observation. However, due to the different installation positions, shooting angles and imaging parameters of different cameras, the appearance of the same target varies in different images. When switching between cameras, it is easy to lose the target or mistakenly associate different targets as the same object, resulting in changes in the tracking number. It is difficult for the on-duty personnel to confirm whether multiple images correspond to the same intrusion target.

[0004] While some linkage solutions can control the camera to rotate according to a preset direction, the timing of the rotation is usually determined based on a fixed time or simple positional relationship. They fail to make a comprehensive judgment by combining the actual speed of the target, the predicted arrival time, and the time taken for the camera to rotate, which may result in problems such as taking over too early, too late, or taking over the wrong target. Furthermore, when a target approaches the boundary of a restricted area, the position of the detection result is prone to change due to frame jitter, occlusion, or short-term positioning error. If an alarm is triggered directly based on the single-frame detection result, it will lead to frequent instantaneous false alarms. Even if the system detects intrusion, the alarm information is often presented in the form of scattered single-frame images or fragments, lacking a continuously traceable cross-frame trajectory, and cannot provide a complete spatiotemporal basis for event review and risk level assessment. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing perimeter monitoring, such as inconsistent spatial coordinates of multiple cameras, misjudgment due to boundary crossing jitter, inaccurate timing of cross-camera takeover, and easy interruption of target identity and trajectory. It provides a perimeter intrusion intelligent identification and tracking method and system with multi-camera linkage, which enables the same intrusion target to take over continuously between adjacent cameras, maintain the same tracking number, and form a cross-camera tracking trajectory that can be used for hierarchical alarms and event backtracking.

[0006] To achieve the above objectives, this solution provides the following four aspects: Firstly, this application provides a multi-camera coordinated perimeter intrusion intelligent recognition and tracking method, including: The system acquires video footage from multiple cameras within the target area, and based on the installation pose calibration relationship of each camera, converts the moving targets detected in each video footage into the actual location information of the target area. The actual location information is compared with the pre-configured perimeter restricted area. After confirming that the moving target has entered the perimeter restricted area, the moving target is identified as an intrusion target and an intrusion trigger record is generated. In response to the intrusion trigger record, the next target camera is selected based on the field of view continuity relationship of each camera and cross-camera takeover is completed by rotating the camera in a preset position. During the cross-camera takeover of the target camera, the identity of the intruding target is maintained based on the target appearance features extracted by the current camera, so as to maintain the same tracking number; By aggregating the observation results of multiple video segments under the same tracking number, a cross-camera tracking trajectory of the intruding target is generated; Based on the cross-camera tracking trajectory, an alarm command is output.

[0007] Optionally, the moving targets detected in each video frame are uniformly converted into the actual location information of the target area, including: Based on the pose calibration relationship of each camera, the landing point of the moving target at the bottom of the screen is converted into the current coordinates of the target area; When the current coordinates calculated by multiple cameras within the same time period are within a preset range, the multiple sets of current coordinates will be merged into the same moving target; The current coordinates output by the camera in the unobstructed view of the moving target are selected as the actual position information of the moving target.

[0008] Optionally, the actual location information is compared with the pre-configured perimeter restricted area. After confirming that the moving target has entered the perimeter restricted area, the moving target is identified as an intrusion target and an intrusion trigger record is generated, including: The actual location information of the same moving target is read sequentially according to the acquisition time, and it is determined whether the actual location is within the perimeter restricted area. When the previous actual position of the moving target is outside the perimeter restricted area and the current actual position is within the perimeter restricted area, the intersection of the line connecting the previous actual position and the current actual position with the boundary of the perimeter restricted area is recorded as the boundary crossing position. Starting from the acquisition time corresponding to the boundary crossing position, continue to read the subsequent actual position of the moving target, and count the number of actual positions that continuously fall into the perimeter restricted area; When the number of actual locations that fall within the perimeter restricted area reaches a set number and the distance of any one of the actual locations from the perimeter restricted area exceeds a set span, the moving target is determined to be an intrusion target. The tracking number of the intrusion target, the boundary crossing location, the boundary crossing time, the current actual location, and the current camera identifier are summarized to generate an intrusion trigger record.

[0009] Optionally, in response to the intrusion trigger record, the next target camera is selected based on the field-of-view continuity relationship of each camera to complete the cross-camera takeover by rotating to a preset position, including: Based on the current actual location and displacement direction recorded in the intrusion trigger record, the expected exit point of the intrusion target from the current camera's field of view is determined; Find out each manageable camera connected to the field of view exit from the pre-configured field of view connection relationship, as well as the target entry area and preset position parameters corresponding to each manageable camera; Based on the current actual location and displacement velocity of the intrusion target, the predicted arrival time of the intrusion target in the target entry area is calculated; The camera whose preset position image covers the position corresponding to the predicted arrival time and whose calibration rotation time is shorter than the predicted arrival time is identified as the target camera. Send a preset position control command to the target camera and detect moving targets in the area where the intrusion target enters, starting from a set time period before the predicted arrival time; If no camera is detected within the preset takeover period, the next takeover camera will be controlled to rotate to the preset position.

[0010] Optionally, during the cross-camera takeover of the target camera, the identity of the intruding target is maintained based on the target appearance features extracted by the current camera to maintain the same tracking number, including: Before the intrusion target leaves the current camera's field of view, extract the combination of the intrusion target's shape features from the unobstructed image of the intrusion target, and use it as the current target's appearance features; After the target camera is rotated into position, the shape feature combination of each moving target to be compared that appears within the perimeter restricted area of ​​the intrusion target is extracted as the appearance feature of the target to be taken over. Filter out moving targets to be compared that appear outside the preset takeover period; The remaining moving targets to be compared are compared with the current target's appearance features one by one. The object with the most matching items is identified as the intrusion target, and the tracking number of the intrusion target is directly used. If two or more objects have the same number of matching items, continue reading the next frame for confirmation.

[0011] Optionally, the observation results of multiple video segments under the same tracking number are aggregated to generate the cross-camera tracking trajectory of the intrusion target, including: Connect the actual positions recorded by the same camera for the same tracking number according to the acquisition time to form a single-lens tracking segment; Connect the last actual position of the previous single-mirror tracking segment directly with the starting actual position of the next single-mirror tracking segment to obtain a cross-mirror continuation segment. Each single-camera tracking segment and each cross-camera connection segment are spliced ​​together in chronological order, and the camera identifiers before and after each camera switch are recorded at the position where the camera switches, so as to obtain the cross-camera tracking trajectory corresponding to the same tracking number.

[0012] Optionally, based on the cross-camera tracking trajectory, an alarm command is output, including: The boundary crossing position, cumulative duration of stay within the restricted area, current location, and current camera identifier of the intruding target are read from the cross-camera tracking trajectory. When the tracking number has not been confirmed in the aforementioned identity preservation steps, the boundary crossing location and current screen information will be pushed to the duty terminal as information to be reviewed. After the tracking number is confirmed to be used, the corresponding alarm level is determined based on the distance between the intrusion target and the key protection location as shown by the cross-camera tracking trajectory. The tracking number, alarm level, and cross-camera tracking trajectory are encapsulated into alarm commands and sent to the duty terminal and on-site alarm device corresponding to the perimeter restricted area. When the location of the intrusion target or the alarm level changes, the alarm command will be reissued according to the updated content.

[0013] Secondly, this application provides a multi-camera linked perimeter intrusion intelligent recognition and tracking system, including: The acquisition module is used to acquire video images captured by multiple cameras within the target area, and based on the installation pose calibration relationship of each camera, convert the moving targets detected in each video image into the actual position information of the target area. The comparison module is used to compare the actual location information with the pre-configured perimeter restricted area. After confirming that the moving target has entered the perimeter restricted area, the moving target is identified as an intrusion target and an intrusion trigger record is generated. The response module is used to respond to the intrusion trigger record and select the next target camera based on the field of view continuity relationship of each camera to complete the cross-camera takeover by rotating the camera in a preset position. The identity retention module is used to retain the identity of the intruding target based on the target appearance features extracted by the current camera during the cross-camera takeover of the target camera, so as to maintain the same tracking number; The aggregation module is used to aggregate the observation results of multiple video segments under the same tracking number and generate the cross-camera tracking trajectory of the intrusion target; An output module is used to output an alarm command based on the cross-camera tracking trajectory. Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be invoked and executed by the processing component to implement a multi-camera coordinated perimeter intrusion intelligent recognition and tracking method as described in the first aspect above.

[0014] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a multi-camera linkage perimeter intrusion intelligent identification and tracking method as described in the first aspect.

[0015] This application converts the pixel points in different camera images to the same target area coordinate system by installing pose calibration relationships, which can merge the same moving target in overlapping fields of view and provide a unified position basis for boundary judgment and cross-camera takeover.

[0016] This application uses boundary crossing, consecutive intrusion count, and inward span to jointly confirm intrusion, which can suppress instantaneous false alarms caused by target box jitter, hovering near the boundary, and short-term positioning errors.

[0017] This application selects the takeover camera based on the continuity of the field of view, the predicted arrival time, the coverage of the preset position, and the rotation time, and sets a candidate switching mechanism after the takeover fails, which can improve the matching degree between the cross-camera takeover timing and the target motion state.

[0018] This application combines the timing constraints with the matching of multiple appearance features, uses the same tracking number after camera switching, and forms a continuous trajectory through single-lens tracking segments and cross-lens transition segments, so that alarm information can be continuously updated as the target position and risk level change.

[0019] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the perimeter intrusion intelligent recognition and tracking system with multi-camera linkage according to the present invention.

[0022] Figure 2 This is an overall flowchart of the multi-camera linkage perimeter intrusion intelligent identification and tracking method of the present invention.

[0023] Figure 3 This is a schematic diagram illustrating the camera calibration, coordinate conversion, and multi-camera position fusion of the present invention.

[0024] Figure 4 This is a network diagram illustrating the continuity of the field of view and the selection of takeover candidates in this invention.

[0025] Figure 5 This is a timing diagram of the cross-camera takeover and identity preservation in this invention.

[0026] Figure 6 This is a schematic diagram illustrating the cross-mirror trajectory formation and alarm status update of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. These embodiments are used to illustrate the technical concept of the present invention and are not intended to limit the scope of protection. Where there is no conflict, the technical features of the various embodiments can be combined with each other.

[0028] The target area in this invention is a monitoring space described by a unified planar coordinate system, which can be the inner side of a park wall, a traffic facility control area, a material storage yard, or other areas requiring perimeter control. The perimeter restricted area is a closed polygon or a set of multiple closed polygons represented by an ordered set of vertices within the target area. Key protection locations are points, lines, or sub-areas within the restricted area that require further hierarchical protection.

[0029] The field-of-view continuity relationship in this invention does not merely represent the physical proximity of cameras, but rather indicates that after a target leaves a specific field-of-view exit of one camera, it can enter the target entry area of ​​another camera within a specified time. A preset position is a pre-calibrated combination of horizontal angle, pitch angle, and focal length for a rotatable camera, and each preset position is simultaneously associated with its ground coverage polygon and the calibrated rotation time.

[0030] like Figure 1 As shown, system 100 includes an acquisition module 110, a comparison module 120, a response module 130, an identity verification module 140, a summarization module 150, an output module 160, and a configuration data storage module 170. Camera group 180 includes cameras 181, 182, and 183; the number of cameras can be expanded according to the size of the target area. A monitoring terminal and on-site alarm device 190 receive alarm commands generated by the output module 160. Each module can be deployed on the same edge computing device, or distributed across cameras, edge nodes, and a central server.

[0031] The acquisition module 110 receives the video stream and timestamps from the camera group 180, and completes moving target detection, single-camera tracking, coordinate conversion, and overlapping field-of-view position fusion. The comparison module 120 determines the actual boundary crossing based on the perimeter restricted area polygon in the configuration data storage module 170. The response module 130 reads the field-of-view continuity relationship and controls the rotatable camera. The identity preservation module 140 compares cross-camera appearance features. The summarization module 150 saves the time-series positions of the same tracking number. The output module 160 generates, updates, and clears alarms. The reason for this setup is to first establish a unified spatial location and reliable intrusion facts, so that subsequent takeover and identity matching have a definite target and time boundary.

[0032] To avoid ambiguity in the procedural terms of claim 1, the following terms are interpreted in conjunction with the data processing procedure of this invention. Each term represents a technical object or processing result in this invention and is not intended to replace the meaning given in this specification with its everyday language meaning.

[0033] The target area is the site area monitored by multiple cameras and established using the same actual length unit to establish a coordinate system.

[0034] The target area is not the rectangular view of any camera, but the actual spatial range determined by the on-site survey boundary, the origin of the coordinate system, and the direction of the coordinate axes; the view position of each camera can be mapped to this range through calibration relationships.

[0035] Moving targets are records of movable objects detected in video footage by the target detection model and linked through consecutive frames from a single camera. A moving target record must include at least the target category, frame bounding box and acquisition time, camera identifier, and single-camera tracking number. Objects with only a single-frame detection box but not yet linked through consecutive frames are not considered detection candidates and are not directly used for boundary crossing confirmation.

[0036] The installation pose calibration relationship is a spatial conversion relationship obtained by using the corresponding data between the control points on the screen and the measured coordinates of the target area after the camera is installed. This term not only refers to the physical installation angle of the camera, but also includes lens distortion parameters, homography matrix or equivalent intrinsic and extrinsic parameters, effective field of view ground polygon, and calibration verification error. Its function is to convert the pixel position of the screen into the coordinates of the target area.

[0037] Actual location information is a record of the position of a moving target in the target area coordinate system after time alignment, spatial merging, and unobstructed observation selection of the current coordinates output by one or more cameras within the same time period. This record includes at least the horizontal coordinate, vertical coordinate, acquisition time, location source camera identifier, and location quality, and is not the uncalibrated image pixel coordinates.

[0038] The perimeter restricted area is one or more pre-configured closed polygons in the target area coordinate system, used to represent the actual area where unauthorized entry by a target is prohibited. Its configuration data consists of sequentially connected boundary vertices and restricted area markers; entering the perimeter restricted area as referred to in this specification means that the actual position information crosses the boundary from the outside of the polygon and meets the conditions of continuous fall and inward span, rather than the target detection box making instantaneous contact with the screen warning line.

[0039] Intrusion targets are moving targets that have been confirmed through boundary crossings, consecutive intrusions, and inward span. Moving targets that do not meet these confirmation criteria are considered candidate intrusion targets and are not considered confirmed intrusion targets. This distinction is used to prevent subsequent camera linkage from being triggered by a single positional jitter.

[0040] The intrusion trigger record is a structured event record generated by the comparison module after confirming the intrusion, rather than a single switch signal. The record contains at least the tracking number, boundary crossing location, boundary crossing time, current actual location, displacement speed, displacement direction, current camera identifier, and restricted area identifier. The response module uses this record as input to select the camera to take over.

[0041] The current camera is the camera that is outputting valid observation records of the intruding target at the start of a cross-camera takeover; the target camera is the next takeover camera selected by the response module based on the field of view continuity, predicted arrival time, preset position coverage, and rotation time; after a successful takeover, the original target camera can become the new current camera in subsequent takeover events. Therefore, the two represent functional roles in the takeover process, rather than fixed equipment categories.

[0042] The field-of-view continuity relationship is a directed correlation data that points from the specific field-of-view exit of the current camera to the area where the target of the takeover camera enters. This relationship also stores target category, typical passage time range, available preset position, ground coverage range, rotation time and candidate priority information, and is not equivalent to the two cameras being physically adjacent.

[0043] The preset position is a set of horizontal angle, pitch angle and focal length control parameters that are pre-calibrated and saved by the rotatable camera, as well as the ground coverage polygon and calibration rotation time corresponding to the set of parameters. Rotating through the preset position means sending the set of parameters to the camera and confirming the position based on the feedback from the device, rather than simply issuing a rotation command in a certain direction without feedback.

[0044] Cross-camera takeover is a continuous process whereby the response module selects the target camera, controls the target camera to switch to a preset position, detects candidate objects during the predicted takeover period, and the identity preservation module confirms the same object before and after the intrusion target leaves the current camera's field of view. Its completion conditions include the target camera being in place, the time of appearance meeting constraints, and the target's identity being associated. It is not a simple video screen switching.

[0045] Target appearance features are a combination of comparable features extracted from unobstructed images of intrusive targets and organized by item, including at least several of the following: target category, aspect ratio and partition color histogram, directional gradient histogram, local texture histogram, and appearance embedding vector; the term refers to technical features used for cross-camera target association and does not refer to the name, identification information, or social identity of a natural person.

[0046] Identity preservation is the process of establishing an observation record association between the unique matching object in the target camera and the intrusion target in the current camera, and making it retain the original tracking number. The tracking number is a logical identifier assigned by the system for a continuous target trajectory. Its purpose is to associate the location, image and alarm data generated by different cameras, and it is not a permanent identity number of a natural person.

[0047] The video observation result is a structured record of the tracked target formed by the camera at a single acquisition moment, which includes at least the tracking number, acquisition time, actual location, camera identifier, target frame, identity status, and location quality; the multi-segment video observation result is a series of consecutive observation sequences formed by one or more cameras with the same tracking number, and is not a direct splicing of the original video files.

[0048] Cross-camera tracking trajectory is a spatiotemporal data sequence formed by splicing together each single-camera tracking segment and cross-camera connection segment according to the acquisition time, using the same tracking number as an index. The single-camera tracking segment consists of the continuous actual positions of the same camera. The cross-camera connection segment connects the last position of the previous single-camera tracking segment with the starting position of the next single-camera tracking segment and records the camera identifiers before and after the switch.

[0049] The alarm command is a structured control message generated and issued by the output module based on the cross-camera tracking trajectory. It includes at least the event identifier, tracking number, alarm level, cross-camera tracking trajectory, current actual location, current camera identifier, version number, and generation time. This command is displayed on the monitoring terminal or executed by the on-site alarm device, and is not just a prompt text displayed on the screen.

[0050] Each processing step of this invention is executed consecutively, with the output of the previous step serving as the input for the next. First, the acquisition module converts and merges the moving target from multiple video feeds into actual location information. This actual location information eliminates the coordinate differences between the camera feeds, allowing the comparison module to compare it with the perimeter restricted area within the same coordinate system. Second, the comparison module confirms a genuine boundary breach and outputs an intrusion trigger record. This record provides the target, location, direction, and time, enabling the response module to query the field-of-view continuity and select the target camera. Third, the target camera's rotation and detection only yield possible candidate objects in space and time, not necessarily proving they are the same as the original intrusion target. Therefore, the identity preservation module continues to use the target's appearance features to confirm the object and retains the tracking number. Subsequently, the same tracking number provides a unique index for summarizing observations from multiple cameras, and the summarization module generates a cross-camera tracking trajectory based on this. Finally, the output module reads the current location, dwell time, and distance to key protected locations from this trajectory, generating and updating alarm commands. This data continuation ensures that each step has a definite source and destination.

[0051] like Figure 3 As shown, installation pose calibration is performed on each camera before the system is put into use. For target areas with approximately flat ground, at least four non-collinear control points are selected in the view of each camera, and the coordinates of each control point in the target area coordinate system are measured. The homography matrix from the pixel plane to the ground plane is obtained using direct linear transformation, and then least squares optimization is performed with the goal of minimizing the reprojection error. If there is a significant elevation difference in the target area, the area is divided into multiple local planes, or the intersection of the ray and the ground is performed using the camera's intrinsic and extrinsic parameters and digital elevation data.

[0052] The camera's installation pose calibration relationship includes lens distortion parameters, homography matrix or equivalent intrinsic and extrinsic parameters, effective field-of-view ground polygon, field-of-view exit, target entry area, and ground coverage polygon for each preset position. After calibration, the position error is calculated using verification points not involved in the calculation. When the average error of the verification points exceeds the allowable value set on site, control points are reselected or recalibrated. The allowable value can be set according to the site dimensions; for example, it can be set to 0.3 meters in a personnel perimeter monitoring scenario.

[0053] The video footage is first synchronized in time and distortion corrected. When the camera supports network time synchronization, the same time server is used to write the acquisition time; when network time synchronization is not supported, the edge node's reception time and a pre-measured transmission delay are used for correction. The same time period of multiple images can be defined as a time window in which the acquisition time difference does not exceed the synchronization tolerance, for example, the synchronization tolerance is set to 100 milliseconds.

[0054] The moving target detection model can employ a target detection network consisting of a convolutional backbone network, a feature pyramid, and a detection head. During model construction, video frames of the target location are acquired under daytime, nighttime, rain / fog, backlighting, and occlusion conditions, and bounding boxes of people, vehicles, or other targets requiring protection are labeled. Samples are grouped by acquisition date or camera type to create training, validation, and test sets, avoiding consecutive frames from entering different sets simultaneously. During training, scale transformations, brightness perturbations, random occlusion, and horizontal flipping are applied to the images, with the weighted sum of bounding box regression loss, target presence loss, and category loss used as the optimization objective. Training stops when the validation set loss no longer decreases within a set number of epochs, and a confidence threshold is determined on an independent test set.

[0055] During online detection, the model outputs the target bounding box and class confidence score, and performs non-maximum suppression on overlapping candidate boxes. For bounding boxes in adjacent frames within the same camera, Kalman scattering is used to predict their positions, and Hungarian matching is performed by combining the bounding box intersection-over-union ratio, center distance, and appearance similarity, thereby forming a single-camera tracking number. In this invention, the detection algorithm is used to obtain stable target positions in the scene; its network layer count can be adjusted with computing power, but the training data source, annotation method, loss target, and threshold determination process should be established as described above.

[0056] The midpoint of the bottom edge of the detection box is used as the landing point of the moving target at the bottom of the image. For people, this landing point approximately corresponds to the contact position between their feet and the ground; for vehicles, it corresponds to the contact position between the bottom edge of the vehicle's bounding box and the road plane. Compared to using the center of the bounding box, the bottom landing point is less affected by the target's height or the vehicle's height, making it suitable for mapping to ground coordinates.

[0057] In the formula, and Indicates the first The pixel coordinates of the bottom landing point of the target in the image from the camera, where Hᵢ represents the homography matrix of the camera. and This indicates the current coordinates in the target region's coordinate system. This represents the homogeneous scaling factor. Dividing the first two terms of the matrix product by the third term yields the planar coordinates in actual length units.

[0058] At this point, each camera can convert the bottom landing point of the target tracked by a single lens into the current coordinates within the target area. However, when the same target is in overlapping fields of view, multiple sets of current coordinates will be generated simultaneously. If these are directly sent to the restricted area for comparison, the same object may be judged repeatedly or contradictory results may occur due to different calibration errors. Therefore, the next step is not to immediately determine intrusion, but to first perform target merging and position selection on the coordinates of multiple cameras in the same time period to form a unique actual position information.

[0059] When the fields of view of multiple cameras overlap, the same target may generate multiple sets of current coordinates at the same time. The acquisition module 110 first filters synchronous observations using a time window, and then establishes candidate associations using a spatial distance threshold. When the Euclidean distance between two sets of current coordinates is not greater than the merging radius, and the target categories are consistent, they are placed in the same candidate cluster. The merging radius is jointly determined by the calibration verification error and the target's movement speed. For example, when the average calibration error is 0.3 meters, the synchronization time window is 0.1 seconds, and the maximum design speed of personnel is 4 meters per second, the merging radius can be set to 1 meter.

[0060] An occlusion-free score is calculated for each camera observation within the candidate cluster. This score is composed of the target bounding box integrity, the degree of overlap with other target bounding boxes, whether the target bounding box touches the image edge, and the detection confidence. An integrity score of 1 is assigned when the target bounding box is complete and does not touch the image edge; the higher the proportion of the target bounding box obscured by other foreground elements, the greater the occlusion penalty. The current coordinates output by the camera with the highest occlusion-free score are selected as the actual position information, thus avoiding the simple averaging of multiple sets of coordinates with correlated errors.

[0061] In the formula, Indicates the first Unobstructedness score for road observation Indicates the confidence level of the detection. This indicates the proportion of the visible area of ​​the target bounding box. This indicates the integrity of the target bounding box before it touches the edge of the screen. This represents continuous tracking stability, with all values ​​normalized to 0 to 1. Weights can be determined through a grid search using labeled occluded samples, selecting the camera with the smallest coordinate error; the values ​​mentioned above represent one feasible setting. If the highest scores are the same, the camera with the smaller calibration verification error is selected.

[0062] After location fusion is completed, each moving target retains only one actual location information record with a timestamp and source camera identifier at each acquisition moment. The perimeter restricted area is also stored in the same target area coordinate system, so this actual location information can be directly geometrically compared with the restricted area boundary. Based on this coordinate unification result, the processing flow enters the perimeter boundary crossing confirmation; if location fusion is not completed first, there is no unique calculation basis for subsequent boundary crossing positions, consecutive fall-in numbers, and inward distances.

[0063] like Figure 2As shown, after obtaining the actual location information, the system enters the perimeter restricted area for comparison. The comparison module 120 uses either the ray method or the number of intersections method to determine whether the current coordinates are located inside the restricted area polygon. The ray method emits a ray from the current coordinates in a fixed direction. If the number of intersections between the ray and the polygon edge is odd, it is determined to be inside; if it is even, it is determined to be outside. When the coordinates fall on the boundary line, the previous inside / outside state is retained and marked as the boundary state to prevent switching back and forth caused by numerical rounding.

[0064] The actual position of the same moving target is read at each acquisition time. When the previous actual position is outside the restricted area and the current actual position is inside the restricted area, the intersection points of the line segments of the two positions with each side line segment of the restricted area boundary are calculated. The first valid intersection point reached along the direction of movement is taken as the boundary crossing position. The time corresponding to the intersection point can be obtained by interpolation according to the linear motion ratio between the previous acquisition time and the current acquisition time. This ensures that subsequent consecutive falls into the statistically consistent range are calculated from the true boundary, rather than from some internal position that may be affected by the sampling interval.

[0065] From the moment of boundary crossing, count the number of consecutive actual locations within the restricted area and calculate the shortest distance from each internal location to each line segment of the restricted area boundary. This shortest distance is called the inward distance, representing the depth of the target after crossing the boundary. The number of lines can be determined based on the sampling frequency and allowable confirmation delay. For example, when the merged locations are output at 10 lines per second, the number of lines can be set to 5. The span can be set to 0.5 meters based on calibration error and false alarm prevention requirements. Intrusion is confirmed when the number of consecutive lines reaches 5 and the inward distance of at least one location is greater than 0.5 meters.

[0066] If, during a period of consecutive falls, a location falls outside the restricted area, but the distance from that location to the boundary is less than the calibration tolerance, it can be ignored once as a jitter point. If two consecutive external locations appear, the current candidate boundary crossing is terminated, and the process waits for a new crossing. The number of ignores and the tolerance are both written to the configuration data storage module 170 and verified by replaying the labeled events. This rule does not replace the consecutive fall conditions in the claims, but rather serves as a preprocessing for location measurement anomalies.

[0067] Upon confirmation of an intrusion, an intrusion trigger record is generated. The record includes at least the intrusion target tracking number, boundary crossing location, boundary crossing time, current actual location, current camera identifier, displacement speed, displacement direction, restricted area identifier, and target category. The intrusion trigger record serves as the determining input for the response module 130, eliminating the need to infer the boundary crossing fact again from the original video footage when selecting the camera to take over.

[0068] In the formula, and These are the actual positions before and after the boundary crossing. and These are the corresponding data collection times. For the location where the boundary is crossed, This is the boundary crossing time. When the two positions coincide, interpolation is not performed, and the system waits for subsequent valid positions.

[0069] The output of the boundary crossing confirmation step does not directly select any adjacent camera, but first generates a complete intrusion trigger record. The response module reads the current actual position, displacement speed, displacement direction, current camera identifier, and boundary crossing time from this record, using these fields to determine which view exit the target will leave from and when it may reach the next target entry area. Therefore, the process only proceeds to view continuation query and cross-camera takeover after the moving target is confirmed as an intrusion target and an intrusion trigger record is generated.

[0070] like Figure 4 As shown, the view continuity relationship can be represented as a directed graph. Each node in the graph consists of a camera identifier, a view exit, or a target entry area; a directed edge indicates that a target can enter the target entry area of ​​a subsequent camera from a specific view exit of the previous camera. Edge attributes include at least the target category, typical travel time range, available preset positions, preset position coverage polygon, rotation time, and candidate priority.

[0071] The construction of field-of-view continuity relationships can be accomplished jointly through site surveying and sample trajectories. First, geometric candidate edges are generated based on the ground polygon of the camera's effective field of view and the channel topology. Then, multi-target passage videos are replayed to statistically analyze the time distribution of targets from the previous exit to the next entry area. Edges are retained when the number of valid samples reaches a set threshold and the passage directions are consistent; if samples are lacking, an initial time range is given based on the channel length and design speed, and updated after manual verification during operation. This ensures that the pre-configured relationships have a clear data source, rather than relying solely on camera numbering.

[0072] The response module 130 estimates the target velocity and displacement direction from several recent actual positions. A uniform Kalman filter can be used, with state variables including lateral position, longitudinal position, lateral velocity, and longitudinal velocity. For each received actual position, prediction is first made according to the sampling interval, and then corrected using the observed position. The intersection of the velocity direction ray and the current effective field of view boundary of the camera is determined as the predicted field of view exit. If the velocity is below a stationary threshold, the exit is determined along the most recent effective displacement direction, and the prediction time window is expanded.

[0073] In the formula, For the first Position and velocity state at any given moment The state transition matrix is ​​constructed based on the sampling interval. For actual location observation, For the observation matrix, and These are process noise and observation noise, respectively. The variance of observation noise can be determined by calibration and verification errors, while the variance of process noise can be determined by the acceleration statistics of the labeled trajectory.

[0074] From the directed graph, all candidate edges connecting the expected field of view exit are identified, yielding each takeover camera and its corresponding target entry area. For each candidate, the predicted arrival time of the target at its current speed to the centerline of the target entry area is calculated; when the path has curves, the median typical travel time recorded using the directed edge is used for correction. If the predicted position falls within the preset position coverage polygon, and the current time plus rotation time and stabilization waiting time is earlier than the predicted arrival time, the preset position meets the takeover conditions.

[0075] Among the candidates meeting the takeover criteria, priority is given to cameras with smaller prediction time errors, larger preset position coverage margins, and shorter rotation times. Preset position coverage margin is the shortest distance from the predicted position to the boundary of the coverage polygon. The candidate score is obtained by weighting the time matching degree, the normalized value of the coverage margin, the reciprocal of the rotation time, and the historical takeover success rate. These weights can be verified using labeled takeover events. The camera with the highest score is selected as the target camera.

[0076] In the formula, For the first The takeover scores of each candidate Scoring is given for the time margin by which the rotation completes earlier than the predicted arrival time. To cover margin scores, The rotation time is scored. This represents the historical takeover success rate. All parameters are normalized to 0 to 1. Candidate ranking is only used to select from those that meet coverage and time constraints; cameras that do not meet these hard constraints will not be included in the takeover process.

[0077] After sending the preset position control command to the target camera, the system reads the current horizontal angle, pitch angle, and focal length fed back by the camera. When the difference between the feedback value and the preset position parameter is less than the device's allowable error and remains stable for a sustained period, the camera is considered to have rotated into position. Starting from a set time period before the predicted arrival time, moving target detection is performed within the target's entry area. The set time period can be twice the standard deviation of the prediction time, and a minimum lead time, such as 1 second, can be set.

[0078] If no moving target meeting the spatiotemporal constraints is detected within the preset takeover period, the candidate is marked as not taken over, and the next takeoverable camera in the sequence is controlled to turn to the corresponding preset position. If all candidates are not taken over, the current tracking number remains in the pending confirmation state, and the last known location and possible exit are sent to the monitoring terminal, without establishing a new identity association out of thin air.

[0079] After completing the preset position rotation and target entry area detection, the response module can only obtain one or more moving targets to be compared that appear within the predicted time and predicted area. Proximity in time and location can narrow down the candidate range, but this cannot be used when multiple people or vehicles appear simultaneously. To prevent irrelevant objects from being added to the original trajectory, the next step involves the identity preservation module reading the target appearance features saved by the current camera and comparing them item by item with the target camera's candidate objects; only when a unique object meets the matching rules does the cross-camera takeover switch to the number retention state.

[0080] like Figure 5 As shown, before the intrusion target leaves the current camera's field of view, the identity preservation module 140 selects several target images from the nearest unobstructed frames. The selection of unobstructed frames is based on the aforementioned unobstructed scoring and requires that the target bounding box height reaches the minimum pixel value required for feature extraction. Scale normalization and color correction are performed on each target image, and then a combination of shape features is formed.

[0081] The shape feature combination can include target category, target bounding box aspect ratio, upper region color histogram, lower region color histogram, directional gradient histogram, local binary texture histogram, and appearance embedding vector. The color histogram is binned and normalized using hue, saturation, and brightness spaces; the directional gradient histogram describes the contour orientation distribution; the local binary texture histogram describes local texture; and the appearance embedding vector is output by the re-identification network. Each feature category constitutes an independent matching term, conforming to the takeover rule of item-by-item comparison and counting of matching terms.

[0082] The re-identification network can be constructed using convolutional networks or visual transformation networks. Training samples consist of manually verified images of the same target from multiple cameras. Images of the same target from different cameras are labeled with the same identity, while images of different targets are labeled with different identities. During training, both identity classification loss and triplet loss are used to reduce the embedding distance for the same identity and increase the embedding distance for different identities. Training, validation, and test sets are divided according to identity. The threshold is determined by the matching accuracy and mismatch rate on the validation set and verified on the target site test set.

[0083] After the target camera rotates into position, the shape feature combinations of each moving target to be compared are extracted within the target entry area. First, objects whose appearance time is earlier than the start of the takeover time window or later than the end are removed. Then, objects whose category does not match the intrusion target category are removed. The remaining objects are compared item by item: a difference in aspect ratio not greater than the corresponding threshold is counted as a match; color, gradient, and texture histogram similarities not less than the corresponding thresholds are each counted as a match; and appearance embedding cosine similarity not less than the corresponding threshold is counted as a match.

[0084] In the formula, This represents the appearance embedding vector of the current target. This represents the appearance embedding vector of the target to be compared. The cosine similarity is represented. The embedding vector is normalized to L2 at output. The color histogram can use the Bach coefficient, and the gradient and texture histograms can use the intersection of their histograms; each threshold is determined by the similarity distribution of the same and different targets in the validation set.

[0085] The system counts the number of matches for each object to be compared. The object with the highest number of matches that also meets the minimum number of matches is identified as the intrusion target and its original tracking number is directly retained. The minimum number of matches can be set to, for example, more than half of all valid matches. If two or more objects have the same number of matches, the system reads the next frame and accumulates the new item-by-item comparison results until a unique object with the highest number of matches is found or the takeover period ends.

[0086] When a target temporarily disappears due to occlusion, a new number is not immediately assigned. Instead, the search area is expanded based on the Kalman scattering prediction within the duration of occlusion. If no unique object is found after the duration of occlusion, the identity status is set to "Unconfirmed". The "Unconfirmed" status allows saving candidate images and spatiotemporal trajectories, but the output module 160 must not directly treat uncertain candidates as the same confirmed target.

[0087] Appearance features are used solely for the technical maintenance of target tracking numbers and are not used to determine the name or other identity attributes of natural persons. To reduce the impact of environmental changes, after a target enters a new camera, the appearance feature template can be updated with a confirmed high-quality image. An upper limit can be set for the update ratio, and the initial template can be retained for subsequent comparisons to avoid feature drift caused by single-frame errors.

[0088] The identity preservation step outputs the confirmed object in the target camera and its original tracking number. This tracking number places the image observation results before and after the takeover under the same data index, thereby solving the problem that different cameras generate their own single-lens numbers and cannot be directly summarized. The next step is to query all observation records using the tracking number, first forming each single-lens tracking segment, and then forming a cross-lens connection segment at the camera switch point; if the identity is still unconfirmed, the candidate record and the pending verification status are retained, and the candidate record is not merged into the confirmed trajectory.

[0089] The summary module 150 stores observation records using tracking numbers as indexes. Each record includes the acquisition time, actual location, camera identifier, identity status, and location quality. Records continuously generated by the same camera are sorted by acquisition time and connected sequentially to form a single-camera tracking segment. To eliminate isolated jitter points, a sliding median filter can be used before connecting the segments. The window length is set according to the location output frequency, for example, taking three consecutive locations.

[0090] After cross-camera takeover is confirmed, the actual position of the last segment of the previous single-camera tracking segment is directly connected to the actual position of the first segment of the next single-camera tracking segment to form a cross-camera continuity segment. The continuity segment is not a frame-by-frame observation between imaginary cameras, but rather represents the spatiotemporal relationship between two confirmed observation segments. Therefore, it is separately marked as a continuity type in the trajectory data, and the camera identifiers before and after the switchover, the takeover confidence level, and the duration of no observation are saved.

[0091] Each single-camera tracking segment and the cross-camera continuation segment are spliced ​​together chronologically to obtain the cross-camera tracking trajectory corresponding to the same tracking number. If the start time of the later single-camera tracking segment is earlier than the end time of the earlier single-camera tracking segment, it indicates that the two cameras are observing overlappingly. In this case, the timestamps are used to align the two cameras, and the actual position with the higher unobstructed score is retained. If the unobserved duration exceeds the maximum continuation duration, the current trajectory is terminated, and the subsequent object is treated as a new trajectory to be confirmed.

[0092] The boundary crossing location and time in the cross-camera tracking trajectory are derived from the intrusion trigger record. The cumulative duration of stay within the restricted area is accumulated based on the time interval within the restricted area in the trajectory. The current location is the latest valid actual location, and the current camera identifier is the camera that provided that location. Therefore, all fields required for alarms are derived from the structured trajectory under the same tracking number, avoiding the need to reassemble inconsistent data during the output stage.

[0093] After the cross-camera tracking trajectory is generated, the boundary crossing location of the intrusion event, the current actual location, the current camera identifier, the cumulative duration of stay within the restricted area, and the camera switching history are all aggregated into the same structure. The risk distance and dwell status required for the alarm level should be calculated from this structure, and cannot be determined solely based on a single frame without considering the trajectory. Therefore, the processing flow enters alarm generation and updating after the trajectory stitching is completed, and alarm version updates are continuously triggered by subsequent position increments of the trajectory.

[0094] like Figure 6 As shown, after the intrusion target is confirmed, the system first enters the intrusion state. If the cross-camera identity has not yet been confirmed, the output module 160 generates verification information, including the tracking number, boundary crossing location, last known location, current candidate image, current camera identifier, and candidate camera identifier, and pushes it to the monitoring terminal. The confirmation result from the monitoring personnel can be used to restore the same number or end the candidate association.

[0095] After confirming the continued use of the tracking number, calculate the shortest distance from the target's current actual location to each key protection location. When the key protection location is a point, Euclidean distance is used; when it is a line, the shortest distance from the point to the line segment is used; when it is a polygon, the distance from the point to the boundary is used if the target is outside the polygon, and 0 is used if the target is inside the polygon. Configure the data storage module 170 to save the mapping rules between distance intervals and alarm levels.

[0096] For example, a Level 1 alarm can be set if a target has entered a general perimeter restricted area but is more than 20 meters away from a key protected location; a Level 2 alarm can be set if the target is between 5 and 20 meters away; and a Level 3 alarm can be set if the target is within 5 meters or has entered a key protected sub-area. The distances mentioned above are only one feasible setting; the actual values ​​will be determined through risk assessment based on site dimensions, personnel response time, and target category. Alarm levels can also be increased based on the cumulative time spent within the restricted area under the same distance conditions, but the level determined by the distance rules must not be lowered.

[0097] The alarm command includes at least the event identifier, tracking number, alarm level, cross-camera tracking trajectory, boundary crossing location, current actual location, current camera identifier, and generation time. The output module 160 sends the alarm command to the monitoring terminal and the on-site alarm device. The on-site alarm device can provide audible and visual prompts, and the monitoring terminal displays the current video, trajectory, and camera switching information.

[0098] When the target location is updated, the current camera is switched, the cumulative dwell time exceeds the set value, or the alarm level changes, an alarm command with an incrementing version number is regenerated. The receiving end only applies the command with the newest version number to avoid network out-of-order overwriting of the new status. After the target continuously leaves the restricted area and the clearance confirmation time is reached, a clearance command is sent and the event trajectory is saved; the clearance confirmation also uses the number of consecutive external locations to avoid frequent alarm activation and deactivation due to boundary jitter.

[0099] After the alarm command is issued, the acquisition module continues to receive new video observations with the same tracking number. The new actual location information sequentially goes through restricted area status updates, necessary subsequent cross-camera takeover, identity verification, and trajectory addition, before the output module compares the old and new locations with the alarm level. This forms a closed loop from video acquisition to alarm update, explaining how the final step uses the aforementioned trajectory and how subsequent observations re-enter the processing chain, rather than interrupting the process after the initial alarm.

[0100] For example, in a rectangular perimeter embodiment, the target area uses a Cartesian coordinate system with meters as the unit, and three rotatable cameras 181, 182, and 183 are set along the passage. The effective fields of view of cameras 181 and 182 partially overlap, and cameras 182 and 183 are connected by a non-overlapping passage. The perimeter restricted area is enclosed by eight ordered vertices, and a key protection polygon is set inside the restricted area.

[0101] Six ground control points are selected for each camera to calculate the homography matrix, with the average error of each verification point not exceeding 0.25 meters. The video sampling rate is 25 frames per second, moving target detection is performed every two frames, and the fused positions are output at 10 lines per second. The synchronization tolerance is set to 100 milliseconds, the overlapping field-of-view coordinate merging radius is set to 0.9 meters, the consecutive falling line count is set to 5 lines, and the inward span is set to 0.5 meters.

[0102] A target moves from the field of view of camera 181 into the restricted area. The previous actual location was outside the restricted area, and the next actual location is inside the restricted area. The system calculates the boundary crossing position and begins to continuously count the occurrences. The inward distance from the fifth internal position to the boundary is 0.62 meters, therefore the comparison module 120 confirms the intrusion and generates an intrusion trigger record. If subsequent actual data changes, the above value is only used as the record for this event and does not reverse the already confirmed boundary crossing position.

[0103] The response module 130 determines the eastern view exit of the camera 181 based on the direction of personnel displacement. Figure 4 From the shown field-of-view sequence, two candidates are obtained: camera 182 and camera 183. It is predicted that the person will arrive at the target entry area of ​​camera 182 in 3.4 seconds. The time required for camera 182 to turn to the corresponding preset position and stabilize is 1.2 seconds. Camera 183 requires a total time of 2.8 seconds and has a smaller coverage margin; therefore, camera 182 is selected. The system begins detection 1 second before the predicted arrival time.

[0104] Two people to be compared appeared in the target area of ​​camera 182. One person appeared during the takeover period, while the other appeared earlier than the start of the takeover period; the latter was eliminated. Five out of the six features of the retained object—aspect ratio, top color, bottom color, orientation gradient, texture, and appearance embedding—met the matching threshold, so the object retained its original tracking number. Subsequently, the end point of the single-lens tracking segment of camera 181 was connected to the start point of the single-lens tracking segment of camera 182, forming a cross-lens continuation segment.

[0105] When the target's distance from the boundary of the key protection polygon decreases from 22 meters to 18 meters, the alarm level is updated from Level 1 to Level 2; if the target continues to enter the key protection polygon, it is updated to Level 3. Each update uses the same event identifier and tracking number, and increases the alarm version number. This embodiment illustrates the logical connection between coordinate fusion, actual boundary crossing confirmation, takeover selection, identity preservation, trajectory stitching, and alarm updates within the same data link.

[0106] The acquisition module 110 may include a video access unit, a target detection unit, a single-camera tracking unit, a coordinate conversion unit, and a position fusion unit. The comparison module 120 may include a region judgment unit, a boundary crossing calculation unit, a continuous fall statistics unit, and a trigger record generation unit. The response module 130 may include an exit prediction unit, a candidate query unit, an arrival time prediction unit, a preset position control unit, and a failure switching unit.

[0107] The identity retention module 140 may include an unobstructed image selection unit, a feature extraction unit, a time period filtering unit, an item-by-item matching unit, and a number continuation unit. The aggregation module 150 may include a single-camera tracking segment generation unit, a cross-camera continuity segment generation unit, and a trajectory stitching unit. The output module 160 may include a verification information generation unit, an alarm classification unit, an instruction encapsulation unit, and an update issuance unit. The above units may be implemented by software programs, or by processors, application-specific integrated circuits, or programmable logic devices.

[0108] The computing device includes a processing component, a storage component, a communication component, and an input / output interface. The processing component may include one or more central processing units, graphics processing units, or neural network processors; the storage component may include volatile memory and non-volatile memory; the communication component connects the camera, the monitoring terminal, and the on-site alarm device. Computer instructions in the storage component are invoked by the processing component and then executed... Figure 2 The sequence shown executes location conversion, intrusion confirmation, cross-camera takeover, identity preservation, trajectory generation, and alarm output.

[0109] Computer storage media can be read-only memory, random access memory, solid-state memory, disk, optical disk, or other non-transitory media capable of storing computer programs. Computer programs can be divided into multiple program modules according to their functions, or they can be deployed in combination; as long as they implement the method of this invention when executed by a computer, they belong to the implementation methods of the computer storage media described in this invention.

[0110] The thresholds and rules in the configuration data storage module 170 are determined before the system is enabled using site-defined data, marked out-of-bounds events, and takeover events. The system retains the configuration version, activation time, verification results, and modification records. During operation, the production configuration must not be automatically overwritten by an unverified single event; when adaptive updates are required, candidate configurations are first generated, and they are only authorized for activation after meeting the false alarm rate, false negative rate, and takeover success rate requirements on the offline replay set.

[0111] Video and trajectory data retention periods can be set according to management requirements. Control commands transmitted between modules can include timestamps, serial numbers, and integrity check values; when the target camera fails to return to its position, the video stream is interrupted, or the coordinate calibration fails, the system generates device abnormality information and retains the last valid trajectory, preventing device abnormalities from being mistakenly interpreted as the target leaving.

[0112] The above description is merely one possible implementation of the present invention. Those skilled in the art can make equivalent adjustments to the detection network structure, feature types, threshold values, number of cameras, and deployment locations without departing from the technical concept of the present invention; these adjustments are still based on the common technical foundations of unified actual location, real boundary crossing confirmation, view continuity takeover, appearance identity preservation, cross-camera trajectory generation, and trajectory-based alarms.

Claims

1. A method for intelligent identification and tracking of perimeter intrusion using multi-camera linkage, characterized in that, include: The system acquires video footage from multiple cameras within the target area, and based on the installation pose calibration relationship of each camera, converts the moving targets detected in each video footage into the actual location information of the target area. The actual location information is compared with the pre-configured perimeter restricted area. After confirming that the moving target has entered the perimeter restricted area, the moving target is identified as an intrusion target and an intrusion trigger record is generated. In response to the intrusion trigger record, the next target camera is selected based on the field of view continuity relationship of each camera and cross-camera takeover is completed by rotating the camera in a preset position. During the cross-camera takeover of the target camera, the identity of the intruding target is maintained based on the target appearance features extracted by the current camera, so as to maintain the same tracking number; By aggregating the observation results of multiple video segments under the same tracking number, a cross-camera tracking trajectory of the intruding target is generated; Based on the cross-camera tracking trajectory, an alarm command is output.

2. The method according to claim 1, characterized in that, The moving targets detected in each video frame are uniformly converted into the actual location information of the target area, including: Based on the pose calibration relationship of each camera, the landing point of the moving target at the bottom of the screen is converted into the current coordinates of the target area; When the current coordinates calculated by multiple cameras within the same time period are within a preset range, the multiple sets of current coordinates will be merged into the same moving target; The current coordinates output by the camera in the unobstructed view of the moving target are selected as the actual position information of the moving target.

3. The method according to claim 1, characterized in that, The actual location information is compared with the pre-configured perimeter restricted area. Upon confirming that the moving target has entered the perimeter restricted area, the moving target is identified as an intrusion target and an intrusion trigger record is generated, including: The actual location information of the same moving target is read sequentially according to the acquisition time, and it is determined whether the actual location is within the perimeter restricted area; When the previous actual position of the moving target is outside the perimeter restricted area and the current actual position is within the perimeter restricted area, the intersection of the line connecting the previous actual position and the current actual position with the boundary of the perimeter restricted area is recorded as the boundary crossing position. Starting from the acquisition time corresponding to the boundary crossing position, continue to read the subsequent actual position of the moving target, and count the number of actual positions that continuously fall into the perimeter restricted area; When the number of actual locations that fall within the perimeter restricted area reaches a set number and the distance of any one of the actual locations from the perimeter restricted area exceeds a set span, the moving target is determined to be an intrusion target. The tracking number of the intrusion target, the boundary crossing location, the boundary crossing time, the current actual location, and the current camera identifier are summarized to generate an intrusion trigger record.

4. The method according to claim 1, characterized in that, In response to the intrusion trigger record, based on the field-of-view continuity relationship of each camera, the next target camera is selected and cross-camera takeover is completed by rotating to a preset position, including: Based on the current actual location and displacement direction recorded in the intrusion trigger record, the expected exit point of the intrusion target from the current camera's field of view is determined; Find out each manageable camera connected to the field of view exit from the pre-configured field of view connection relationship, as well as the target entry area and preset position parameters corresponding to each manageable camera; Based on the current actual location and displacement velocity of the intrusion target, the predicted arrival time of the intrusion target in the target entry area is calculated; The camera whose preset position image covers the position corresponding to the predicted arrival time and whose calibration rotation time is shorter than the predicted arrival time is identified as the target camera. Send a preset position control command to the target camera and detect moving targets in the area where the intrusion target enters, starting from a set time period before the predicted arrival time; If no camera is detected within the preset takeover period, the next takeover camera will be controlled to rotate to the preset position.

5. The method according to claim 1, characterized in that, During the cross-camera takeover of the target camera, the identity of the intruding target is maintained based on the target appearance features extracted by the current camera to maintain the same tracking number, including: Before the intrusion target leaves the current camera's field of view, extract the combination of the intrusion target's shape features from the unobstructed image of the intrusion target, and use it as the current target's appearance features; After the target camera is rotated into position, the shape feature combination of each moving target to be compared that appears within the perimeter restricted area of ​​the intrusion target is extracted as the appearance feature of the target to be taken over. Filter out moving targets to be compared that appear outside the preset takeover period; The remaining moving targets to be compared are compared with the current target's appearance features one by one. The object with the most matching items is identified as the intrusion target, and the tracking number of the intrusion target is directly used. If two or more objects have the same number of matching items, continue reading the next frame for confirmation.

6. The method according to claim 1, characterized in that, By aggregating observation results from multiple video clips under the same tracking number, a cross-camera tracking trajectory for the intruding target is generated, including: Connect the actual positions recorded by the same camera for the same tracking number according to the acquisition time to form a single-lens tracking segment; Connect the last actual position of the previous single-mirror tracking segment directly with the starting actual position of the next single-mirror tracking segment to obtain a cross-mirror continuation segment. Each single-camera tracking segment and each cross-camera connection segment are spliced ​​together in chronological order, and the camera identifiers before and after each camera switch are recorded at the position where the camera switches, so as to obtain the cross-camera tracking trajectory corresponding to the same tracking number.

7. The method according to claim 1, characterized in that, Based on the cross-camera tracking trajectory, an alarm command is output, including: The boundary crossing position, cumulative duration of stay within the restricted area, current location, and current camera identifier of the intruding target are read from the cross-camera tracking trajectory. When the tracking number has not been confirmed in the aforementioned identity preservation steps, the boundary crossing location and current screen information will be pushed to the duty terminal as information to be reviewed. After the tracking number is confirmed to be used, the corresponding alarm level is determined based on the distance between the intrusion target and the key protection location as shown by the cross-camera tracking trajectory. The tracking number, alarm level, and cross-camera tracking trajectory are encapsulated into alarm commands and sent to the duty terminal and on-site alarm device corresponding to the perimeter restricted area. When the location of the intrusion target or the alarm level changes, the alarm command will be reissued according to the updated content.

8. A multi-camera linked perimeter intrusion intelligent recognition and tracking system, characterized in that, The system is used to implement the multi-camera linkage perimeter intrusion intelligent identification and tracking method according to any one of claims 1-7; the system includes: The acquisition module is used to acquire video images captured by multiple cameras within the target area, and based on the installation pose calibration relationship of each camera, convert the moving targets detected in each video image into the actual position information of the target area. The comparison module is used to compare the actual location information with the pre-configured perimeter restricted area. After confirming that the moving target has entered the perimeter restricted area, the moving target is identified as an intrusion target and an intrusion trigger record is generated. The response module is used to respond to the intrusion trigger record and select the next target camera based on the field of view continuity relationship of each camera to complete the cross-camera takeover by rotating the camera in a preset position. The identity retention module is used to retain the identity of the intruding target based on the target appearance features extracted by the current camera during the cross-camera takeover of the target camera, so as to maintain the same tracking number; The aggregation module is used to aggregate the observation results of multiple video segments under the same tracking number and generate the cross-camera tracking trajectory of the intrusion target; The output module is used to output alarm commands based on the cross-mirror tracking trajectory.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a multi-camera linkage perimeter intrusion intelligent identification and tracking method as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The device stores a computer program, which, when executed by a computer, implements a multi-camera linkage perimeter intrusion intelligent identification and tracking method as described in any one of claims 1-7.

Images