AI perimeter alarm system integrating vibration optical fiber, radar perception and video monitoring

Abstract

The invention provides an AI perimeter alarm system fusing vibration optical fiber, radar perception and video monitoring, and relates to a perimeter alarm system, the system comprises a data acquisition module respectively obtaining vibration signal data, target point cloud data and video stream data; the space grid mapping module is used for constructing a three-dimensional space grid and generating gridding multi-modal data under a unified space-time reference; the event kernel identification module is used for identifying the grid units of which the vibration intensity values and the target speed values both exceed respective preset thresholds, and determining the grid units as initial event kernels; the threat propagation field generation module is used for generating a threat propagation field on the three-dimensional space grid; the multi-modal consistency scoring module is used for calculating to obtain a multi-modal consistency score; the perimeter alarm decision module is used for judging whether a target activity area formed by grid units which are continuous in space and have multi-modal consistency scores higher than a preset score threshold exists in the threat propagation field or not.

Description

Technical Field

[0001] This application relates to the field of perimeter alarm system technology, and in particular to an AI perimeter alarm system that integrates vibration fiber optics, radar sensing and video surveillance. Background Technology

[0002] Perimeter security systems serve as the first line of defense for critical infrastructure, military bases, important industrial parks, airports, and border areas. Traditional perimeter security technologies primarily rely on single types of sensors, such as infrared beam sensors, electronic fences, leaky cables, or vibrating fiber optics. Vibrating fiber optic sensors are widely used in modern perimeter security due to their high sensitivity, resistance to electromagnetic interference, and ability to be deployed over long distances. They detect intrusion behaviors such as climbing, cutting, and impact by detecting changes in the light signal caused by vibration of optical fibers attached to fences or walls.

[0003] However, systems relying solely on vibrating fiber optic cables have an inherent and insurmountable drawback: a high false alarm rate. Due to their high sensitivity, environmental factors such as strong winds, heavy rain, hail, or even the touch of small animals (like birds or cats) can trigger vibrations and cause false alarms. Operators need to spend a significant amount of time identifying these false alarms, wasting human resources and potentially leading to complacency regarding genuine alarms due to the "boy who cried wolf" effect.

[0004] To address this issue, the industry has begun to explore the use of video surveillance for verification. When a vibration alarm is triggered on the fiber optic cable, the system automatically retrieves footage from the corresponding security zone for security personnel to review manually. This "sensor + video" approach improves accuracy to some extent, but it still heavily relies on human intervention, has a slow response time, and its reliability drops sharply under poor visual conditions such as at night or in inclement weather.

[0005] In recent years, with the maturation and cost reduction of radar technology, especially millimeter-wave radar, it has been introduced into the field of perimeter security. Radar can accurately detect the distance, speed, and angle of targets around the clock, unaffected by light, smoke, rain, or snow, thus compensating for the shortcomings of video surveillance. This has spurred the development of multi-sensor fusion security solutions, which attempt to combine data from vibration fiber optics, radar, and video for comprehensive judgment.

[0006] While multi-sensor fusion is the right direction for development, most existing fusion technologies remain at a relatively superficial level. For example, simple AND logic fusion only triggers an alarm when vibration fiber optics, radar, and video simultaneously detect anomalies in the same area. While this approach can reduce false alarms, it is also highly susceptible to missed detections because, in the initial stages of an intrusion, the timing and spatial location of signals captured by different sensors may vary slightly. More advanced solutions might weight and score alarm signals from different sensors, but this scoring is often based on isolated, instantaneous events, lacking a holistic understanding and modeling of intrusion behavior as a continuous spatiotemporal process. They struggle to distinguish between a harmless individual briefly lingering near a fence and an intruder searching for weaknesses and climbing over it.

[0007] Therefore, a core technical problem currently facing the perimeter security field is: how to deeply and organically integrate the information from three sensors—vibration fiber optics, radar, and video—which have completely different physical principles, data formats, and spatiotemporal characteristics, to construct an intelligent analysis framework that can understand the spatiotemporal continuity and multimodal intrinsic correlation of intrusion behavior. This would fundamentally achieve high-precision identification of real intrusion threats and strong robust suppression of environmental and harmless interference, that is, to achieve "high accuracy and low false alarm" in perimeter alarms. Summary of the Invention

[0008] This application provides an AI perimeter alarm system that integrates vibration fiber optics, radar sensing, and video surveillance, aiming to at least improve one of the aforementioned technical problems.

[0009] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: This application provides an AI perimeter alarm system integrating vibration fiber optic sensing, radar sensing, and video surveillance. The system includes: a data acquisition module for acquiring vibration signal data output by a vibration fiber optic sensor, target point cloud data output by a radar sensor, and video stream data output by a video surveillance device within a preset perimeter defense zone; a spatial grid mapping module for constructing a three-dimensional spatial grid based on the physical dimensions of the preset perimeter defense zone; the spatial grid mapping module also maps the vibration signal data, the target point cloud data, and the video stream data to corresponding grid cells of the three-dimensional spatial grid to generate gridded multimodal data under a unified spatiotemporal reference; and an event kernel recognition module for identifying grid cells in the gridded multimodal data where both the vibration intensity value and the target velocity value exceed their respective preset thresholds. The initial event kernel is determined from the grid cell. The threat propagation field generation module generates a threat propagation field on the three-dimensional spatial grid, starting from the initial event kernel and based on the spatial adjacency between the grid cells and the target motion continuity indicated by the target point cloud data. The multimodal consistency scoring module calculates a multimodal consistency score for each grid cell within the coverage area of ​​the threat propagation field by fusing the vibration intensity value it carries, the target point cloud data attributes, and the target semantic category in the video stream data. The perimeter alarm decision module determines whether there exists a target activity area within the threat propagation field consisting of spatially continuous grid cells whose multimodal consistency scores are all higher than a preset score threshold, and outputs a perimeter intrusion alarm signal when the determination is yes.

[0010] In one possible implementation, the spatial grid mapping module is specifically used to: divide the three-dimensional spatial grid into multiple grid cells along the length and height directions of the perimeter defense zone; distribute the vibration signal data along the length direction of the perimeter defense zone to each grid cell adjacent to the perimeter boundary, and assign a vibration intensity value to each grid cell; place each target point in the target point cloud data into a unique grid cell according to its three-dimensional spatial coordinates, and assign the target point's velocity and radar cross section attributes to the grid cell; perform target recognition on the video stream data, obtain the target's semantic category and two-dimensional image position, and convert the two-dimensional image position into three-dimensional spatial coordinates by combining depth information, and then assign the semantic category to the grid cell corresponding to its coordinates.

[0011] In one possible implementation, the threat propagation field generation module is specifically used to: mark the initial event kernel as active; starting from the initial event kernel, iteratively check the neighboring grid cells of all grid cells in the active state; if the target point cloud data contained in a neighboring grid cell shows that it is continuous with the target point cloud data in the currently active grid cell in the motion trajectory, then mark the neighboring grid cell as active. The set of all grid cells marked as active is defined as the threat propagation field.

[0012] In one possible implementation, the continuity of the motion trajectory means that the difference in direction between the velocity vector of the target point in the adjacent grid cell and the velocity vector of the target point in the currently active grid cell is less than a preset angle and the rate of change of velocity magnitude is within a preset range.

[0013] In one possible implementation, the multimodal consistency scoring module is specifically used to: construct a multidimensional feature vector for each grid cell based on its vibration intensity value, target velocity value, target radar cross section value, and target semantic category; input the multidimensional feature vector into a preset scoring function, wherein the scoring function combines the feature values ​​of each dimension with preset weights to output the multimodal consistency score.

[0014] In one possible implementation, the target semantic categories include “person”, “vehicle”, “animal”, and “environmental disturbance”; the scoring function assigns a higher weight to the “person” semantic category than to the other semantic categories.

[0015] In one possible implementation, the target active region is a connected subset in a three-dimensional spatial grid whose volume is greater than a preset volume threshold, and whose multimodal consistency scores of all internal grid cells are higher than the preset score threshold.

[0016] In one possible implementation, the system further includes: an environmental baseline adaptation module, used to statistically analyze the distribution of the vibration intensity value and the target point cloud data attributes over multiple historical time periods in which no alarm signal is generated, in order to construct an environmental noise baseline model; the environmental baseline adaptation module is also used to dynamically adjust the preset threshold values ​​for vibration intensity value and target velocity value used by the event kernel identification module according to the environmental noise baseline model.

[0017] In one possible implementation, the environmental noise baseline model records the vibration intensity values ​​and statistical characteristics of the target point cloud data attributes under different weather conditions or at different times of the day; the environmental baseline adaptive module selects the corresponding statistical characteristics to adjust the preset threshold based on the current time and weather information.

[0018] In one possible implementation, the system further includes a threat level assessment module, which, after outputting the perimeter intrusion alarm signal, comprehensively assesses and outputs the threat level of the intrusion event based on the volume of the target activity area, the average multimodal consistency score of its internal grid cells, and the target semantic category. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of an AI perimeter alarm system that integrates vibration fiber optics, radar sensing, and video surveillance, provided for some embodiments of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0021] In the following description, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0022] Furthermore, in this application, directional terms such as "upper," "lower," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and may change accordingly depending on the orientation of the components in the accompanying drawings.

[0023] In this application, unless otherwise expressly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium. Furthermore, the term "electrical connection" can refer to the manner in which an electrical connection is used to achieve signal transmission.

[0024] As used herein, “about,” “approximately,” or “approximately” includes the stated value and a reference value within an acceptable range of deviation from the given value, characterized in that the acceptable range of deviation is determined by a person skilled in the art taking into account the measurement under discussion and the error associated with the measurement of the given quantity (i.e., the limitations of the measurement method).

[0025] This application provides an AI perimeter alarm system integrating vibration fiber optic sensing, radar perception, and video surveillance. This system aims to address the problems of low accuracy and high false alarm rates in identifying genuine intrusion behavior in existing perimeter security technologies, caused by insufficient information dimensionality from single sensors or inadequate fusion of multi-sensor information. This application achieves accurate judgment of perimeter intrusion events by unifying sensor data with different physical properties into a high-dimensional spatiotemporal grid and establishing an event propagation and consistency verification mechanism based on this grid.

[0026] The architecture and workflow of the system in one embodiment of this application will be described in detail below. For example... Figure 1 As shown, the system architecture may include multiple functional modules, which may be a combination of hardware, software, or firmware, for performing the specific functions described above.

[0027] The system may include a data acquisition module, a spatial grid mapping module, an event kernel identification module, a threat propagation field generation module, a multimodal consistency scoring module, and a perimeter alarm decision module.

[0028] The data acquisition module is the data input source for the entire system. Its core task is to acquire raw sensing data synchronously or asynchronously from multiple heterogeneous sensors deployed within the perimeter defense zone. In this embodiment, the data acquisition module integrates access interfaces for three core sensors.

[0029] The data acquisition module acquires vibration signal data output from the fiber optic vibration sensor. Fiber optic vibration sensors are typically installed along physical fences, walls, or underground, and are extremely sensitive to mechanical disturbances occurring at the perimeter. Their raw output data is a one-dimensional time-series signal, where the amplitude, frequency, and mode reflect the intensity and nature of the vibration event. The data acquisition module segments and preprocesses this time-series signal, for example, by using short-time Fourier transform or wavelet analysis to extract the vibration energy or characteristic frequencies at different locations along the fiber within each time window, ultimately forming vibration signal data associated with the physical location of the fiber.

[0030] The data acquisition module also acquires target point cloud data output by the radar sensors. Millimeter-wave radar or lidar is deployed at key locations capable of covering the entire perimeter defense zone. The radar can detect moving targets within the defense zone around the clock, unaffected by lighting conditions or inclement weather. Its output is a series of target point clouds, each point typically containing its position information in a three-dimensional Cartesian coordinate system. And Doppler velocity information relative to radar Some advanced radars can also provide the radar cross section (RCS) value of the target. This value reflects, to some extent, the size and material of the target. The data acquisition module is responsible for aggregating data from one or more radars and aligning them in a coordinate system to form a target point cloud set in a unified coordinate system.

[0031] The data acquisition module also acquires video stream data output by the video surveillance equipment. High-definition cameras are also deployed along the perimeter, providing rich visual information. The data acquisition module receives real-time video streams, which are continuous sequences of image frames.

[0032] Understandably, to ensure data consistency, the data acquisition module assigns a timestamp to all collected data. .

[0033] To break down the barriers between different sensor data in terms of structure, dimension, and physical meaning, this application constructs a three-dimensional spatial mesh as a unified analysis framework. The spatial mesh mapping module is responsible for creating this mesh and projecting the acquired multi-source heterogeneous data into this shared representation space.

[0034] This module first constructs a three-dimensional spatial mesh based on the physical dimensions of the preset perimeter defense zone. For example, suppose the perimeter defense zone is a fence 100 meters long and 3 meters high, and the depth to be monitored is 5 meters. Then, a three-dimensional space with dimensions of 100 meters × 5 meters × 3 meters (length × width × height) can be constructed. For discretization analysis, this space is divided into a series of small, regular cubic units, i.e., grid cells. ,in This is the 3D index of the cell within the mesh. The size of each mesh cell, for example... The spatial resolution of the system is determined by the spatial parameters. Higher resolution results in more accurate positioning, but also in greater computational overhead.

[0035] Subsequently, the spatial grid mapping module performs the mapping process for multimodal data. This process includes vibration data mapping, radar data mapping, and video data mapping.

[0036] Vibration data mapping distributes one-dimensional vibration signal data into a three-dimensional mesh. Since the vibrating fiber is linearly deployed, its location is deterministic. Therefore, the module precisely maps each segment of the vibrating fiber to a row of mesh cells immediately adjacent to the physical perimeter (e.g., a fence). At a certain timestamp... If the fiber's first The segment detected a vibration energy value. This value will then be assigned to or accumulated in the corresponding grid cell. (in The vibration intensity value that forms the grid cell is based on a specific property of the layer (representing the layer immediately adjacent to the fence). .

[0037] Radar data mapping then fills discrete target point cloud data into a three-dimensional mesh. For timestamps... Each radar target point collected The module first calculates the index of its corresponding grid cell. The calculation method is as follows: , , ,in This refers to the dimensions of the grid cell in each dimension. Then, this grid cell... It is marked as "occupied" and assigned attributes such as speed to the target point. and RCS value If multiple points fall into the same cell, attributes can be aggregated, for example, by taking the average velocity or the maximum RCS value. In this way, the originally sparse point cloud data is transformed into a sparse field representing target occupancy, motion, and physical characteristics on a grid.

[0038] Video data mapping is responsible for imbuing three-dimensional space with high-level semantic information from two-dimensional images. This module can integrate an existing target recognition component (e.g., a deep learning-based YOLO or Faster R-CNN model) that processes video frames, identifies targets in the image, and assigns their categories (e.g., "person," "vehicle," "animal," or potentially distracting "floating plastic bag") and two-dimensional bounding boxes. The spatial mesh mapping module utilizes camera calibration parameters and depth estimation algorithms (which can be based on radar data, for example, associating the center of the target identified in the video with the nearest radar point cloud cluster in the same direction to directly obtain its depth information) to convert the center point of the two-dimensional bounding box into three-dimensional spatial coordinates. Then, the grid cell to which that coordinate belongs is calculated. and the identified semantic categories Assign this unit.

[0039] After processing by the spatial grid mapping module, the system obtains a value for each timestamp. A dynamically updated, information-rich gridded multimodal data structure. Each grid cell... Both may include vibration intensity values. Target point cloud attributes (occupancy status) velocity vector RCS and target semantic category .

[0040] After obtaining unified gridded multimodal data, the system needs to quickly pinpoint potential, high-threat intrusion initiation points from massive amounts of background noise and harmless activities.

[0041] A real, threatening intrusion typically triggers both a violent physical disturbance and the approaching movement of the target at the instant of physical contact with the perimeter. Therefore, the event kernel identification module continuously scans all mesh cells within a time window, searching for cells that meet specific dual conditions.

[0042] Specifically, the event kernel identification module will inspect each grid cell. The properties of the mesh element it seeks. The mesh element it searches for must simultaneously satisfy two conditions: Its vibration intensity value Exceeding a preset vibration threshold This threshold is used to filter out low-intensity vibrations caused by a breeze, a small animal bumping into you, etc.

[0043] The magnitude of the velocity value in the target point cloud data attributes it carries. Exceeding a preset speed threshold This threshold is designed to confirm the presence of a moving entity at that location, rather than a static object or sensor noise.

[0044] For example, suppose It is set to 80 (on a normalized scale of 0-100). It is set to 0.2 m / s. If the grid cells In timestamp of If the value is 85 and its internal radar point display speed is 0.5 m / s, then the module will... It was identified.

[0045] One or more grid cells that meet the above conditions and are identified within the same time window are defined by the system as the "initial event kernel". The initial event kernel represents the spacetime singularity where the intrusion is most likely to occur.

[0046] Identifying an isolated initial event kernel is insufficient to constitute a complete intrusion event assessment. An intelligent intruder is a continuously moving entity. Therefore, this application also includes a threat propagation field generation module to track and depict the complete activity trajectory and scope of influence associated with the initial event kernel on a three-dimensional spatial grid.

[0047] Threat Propagation Field The generation process is designed as a deterministic expansion process based on spatial adjacency and motion continuity. This process can be viewed as a directed search on a mesh graph.

[0048] This module first identifies all initial event kernels. A cell is added to a "pending" queue and its state in the mesh is marked as "active". The module then enters an iterative loop. In each iteration, it retrieves an active mesh cell from the pending queue. And check all adjacent grid cells in its space. (For example, in a 3D mesh, a cell has 26 neighboring cells.)

[0049] For each unchecked adjacent grid cell The module will then perform a motion continuity check. This check aims to confirm... Are the activity goals in the middle the same as those in the middle? The reasonable continuation of the target. In this embodiment, motion continuity is determined by comparing the target velocity vectors within two cells. Specifically, if Includes target point cloud data, and its velocity vector. and velocity vector A property is considered to have continuity if it meets the following conditions: The angle between the two velocity vectors is less than a preset angle threshold. For example, 30 degrees. This indicates that the target did not undergo a physically unreasonable sudden change.

[0050] Rate of change of the magnitudes of the two velocity vectors Within a pre-defined reasonable range, for example This rules out the possibility of a sudden change in the target velocity.

[0051] If an adjacent grid cell judged as being with If it has motion continuity and has not yet been activated, then the module will... The status is also marked as "active" and added to the "pending" queue.

[0052] This iterative process continues until the "pending" queue is empty. At this point, the set of all grid cells marked as "active" constitutes the threat propagation field of this event. This field spatially depicts a continuous, dynamic path and range of target activity, starting from the initial point of contact. For example, if a person climbs over a fence (generating the initial event kernel) and then moves on the ground, the threat propagation field will start from a point on the fence and form an "activated" trajectory in a three-dimensional grid along the path the person has taken.

[0053] Threat propagation fields depict the trajectory of activity but do not definitively confirm its nature. For example, a large dog jumping onto and off a fence could create a similar propagation field. To distinguish between real and non-threat activities, a multimodal consistency scoring module was introduced to quantitatively assess the "threat attributes" of each point within the propagation field.

[0054] The goal of this module is to target threat propagation environments. Each grid cell within the coverage area Calculate a comprehensive multimodal consistency score. The higher this score, the more consistently the multi-source information carried by the unit points to a high-threat event.

[0055] In order to calculate This module first constructs a multidimensional feature vector for each relevant grid cell. This vector aggregates key information from different sensors. In this embodiment, the feature vector... Includes the following dimensions: Normalized vibration intensity value Normalized target velocity magnitude Normalized target radar cross section value and a discrete code representing the target semantic category. .

[0056] For example, the target semantic category identified in the video is encoded as a numerical value: Represents "person" Represents "vehicles" Represents "animals" This represents "environmental interference" or "unknown".

[0057] Then, the module will convert the feature vector Input a preset scoring function To calculate This function is a deterministic combinatorial operation designed to reward patterns where all metrics point to a high threat level. A simple implementation is a weighted summation: in, These are preset weights that reflect the importance of different features in threat assessment. For example, (Semantic category weight) and (Vibration intensity weight) may be assigned a high value because "high-intensity vibrations" induced by "human" targets on the perimeter are the most typical intrusion feature.

[0058] Through this process, each grid cell in the threat propagation field is assigned a quantified threat score. An event caused by a "person," resulting in violent vibrations and continuous movement on the fence, assigns a threat score to all grid cells along its path. The score will be significantly higher than if a kitten were to pass by silently.

[0059] The perimeter alarm decision module receives the threat propagation field. and the multimodal consistency score of each grid cell. As input, its decision-making logic is based on the search and verification of a "target activity area".

[0060] The module can be set with a preset score threshold. For example, 0.8 (in the range of 0 to 1). It is in the threat propagation field. Filter out all Value higher than The module then performs connectivity analysis on these high-scoring grid cells (e.g., using breadth-first search or depth-first search algorithms) to identify one or more spatially contiguous subsets composed of these cells. Each such connected subset is defined as a potential "target activity region". .

[0061] The module identifies each Verification is performed. The verification criterion is whether the size of the area is sufficient to pose a threat. In this embodiment, this criterion is measured by its volume. Module calculation. The volume is calculated by multiplying the number of grid cells it contains by the volume of a single cell. If this volume exceeds a preset volume threshold... (For example, If the size of the target is roughly the size of an adult, then the system determines that a real, continuous intrusion event caused by a high-threat target has occurred.

[0062] The perimeter alarm decision module will output a high-confidence perimeter intrusion alarm signal only if at least one "target activity area" meets the above conditions. This signal can be used to trigger warning lights and sirens, send alarm notifications to the security center, and can be accompanied by the location of the intrusion (i.e., the location of the intrusion). Detailed information such as coordinates in the grid and the semantic category of the target.

[0063] In some embodiments, the system further includes an environmental baseline adaptation module and a threat level assessment module. The perimeter environment is not static; strong winds, heavy rains, and other weather conditions can cause widespread vibrations in the fence, potentially leading to... Thresholds were frequently triggered falsely. The environmental baseline adaptation module is designed to allow the system to learn and adapt to these environmental changes. During "peacetime" periods when the system is operating normally and no alarms are generated, this module continuously collects and analyzes vibration intensity values. It also tracks the velocity distribution of background clutter detected by radar (such as swaying leaves). It constructs an environmental noise baseline model that records the statistical characteristics (such as mean and variance) of noise data under different conditions, such as "sunny day," "rainy day," and "windy day," or at different times of day, such as "daytime" and "nighttime." When the system is running, it can select the best-matching noise from the baseline model based on the current time and information from external weather sensors, and use this to dynamically and finely adjust the noise used by the event kernel recognition module. and Threshold. For example, on a windy day, It will be moderately enhanced to ignore background wind noise. This adaptive capability significantly reduces false alarms caused by environmental factors.

[0064] After confirming an intrusion and issuing an alert, the threat level assessment module can quantitatively evaluate the severity of the incident. The assessment is based on multiple factors, such as the identified "target activity area." The volume (reflecting the size of the target), the average value of its internal mesh cells The module inputs a score (reflecting the consistency of threat characteristics) and the semantic category of the target identified in the video ("armed personnel" clearly pose a higher threat level than "ordinary personnel"). This input is then fed into a pre-defined evaluation rule or function, outputting a specific threat level, such as "Level 1 (highest)," "Level 2," or "Level 3." This level information helps security personnel allocate response resources more effectively, prioritizing the most serious threats.

[0065] The technical solution of this application constructs a unified three-dimensional spatial grid, deeply aligning and correlating data from three heterogeneous sensors—vibration fiber optics, radar, and video—in a spatiotemporal manner. Instead of relying on simple threshold judgments or isolated data fusion, it creatively proposes a chain-like deterministic reasoning process: "event kernel identification - threat propagation field generation - multimodal consistency scoring." This process characterizes the logic of an intrusion event from "contact point" to "continuous activity trajectory" and then to "nature confirmation," thereby enabling the separation of intrusion behavior from complex backgrounds. Therefore, this application can significantly improve the accuracy of perimeter alarms, greatly reduce false alarm rates, and provide reliable technical protection for the security of critical facilities.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims. Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0067] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another apparatus, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0068] The units described as separate components may or may not be physically separate. A component shown as a unit can be one physical unit or multiple physical units; that is, it can be located in one place or distributed in multiple different places. Depending on actual needs, some or all of the units can be selected to achieve the purpose of this embodiment.

[0069] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware.

[0070] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI perimeter alarm system integrating vibration fiber optics, radar sensing, and video surveillance, characterized in that, The system includes: The data acquisition module is used to acquire vibration signal data output by the vibration fiber optic sensor, target point cloud data output by the radar sensor, and video stream data output by the video monitoring equipment within the preset perimeter defense zone. The spatial grid mapping module is used to construct a three-dimensional spatial grid based on the physical dimensions of the preset perimeter defense zone; the spatial grid mapping module also maps the vibration signal data, the target point cloud data and the video stream data to the corresponding grid cells of the three-dimensional spatial grid to generate gridded multimodal data under a unified spatiotemporal reference. The event kernel identification module is used to identify grid cells in the gridded multimodal data where both the vibration intensity value and the target velocity value exceed their respective preset thresholds, and to determine the grid cell as the initial event kernel; The threat propagation field generation module is used to generate a threat propagation field on the three-dimensional spatial grid, starting from the initial event kernel and based on the spatial adjacency relationship between the grid cells and the target motion continuity indicated by the target point cloud data. The multimodal consistency scoring module is used to calculate a multimodal consistency score for each grid cell within the coverage area of ​​the threat propagation field by fusing the vibration intensity value it carries, the target point cloud data attributes, and the target semantic category in the video stream data. The perimeter alarm decision module is used to determine whether there is a target activity area in the threat propagation field that is composed of spatially continuous grid cells whose multimodal consistency scores are all higher than a preset score threshold, and outputs a perimeter intrusion alarm signal when the determination is yes.

2. The system according to claim 1, characterized in that, The spatial grid mapping module is specifically used for: The three-dimensional spatial grid is divided into multiple grid cells along the length and height of the perimeter defense zone; The vibration signal data is distributed along the length of the perimeter defense zone to each grid cell adjacent to the perimeter boundary, and a vibration intensity value is assigned to each grid cell. Each target point in the target point cloud data is placed into a unique grid cell according to its three-dimensional spatial coordinates, and the grid cell is assigned the velocity and radar cross section attributes of the target point. The video stream data is used to perform target recognition, obtain the semantic category and two-dimensional image position of the target, and convert the two-dimensional image position into three-dimensional spatial coordinates by combining the depth information. Then, the semantic category is assigned to the grid cell corresponding to its coordinates.

3. The system according to claim 1, characterized in that, The threat propagation field generation module is specifically used for: Mark the initial event kernel as active; Starting from the initial event kernel, iteratively check the neighboring grid cells of all active grid cells; If the target point cloud data contained in an adjacent grid cell shows that it is continuous with the target point cloud data in the currently active grid cell in terms of motion trajectory, then the adjacent grid cell is marked as active. The set of all grid cells marked as active is defined as the threat propagation field.

4. The system according to claim 3, characterized in that, The continuity of the motion trajectory means that the difference in direction between the velocity vector of the target point in the adjacent grid cell and the velocity vector of the target point in the currently active grid cell is less than a preset angle and the rate of change of velocity magnitude is within a preset range.

5. The system according to claim 1, characterized in that, The multimodal consistency scoring module is specifically used for: For each grid cell, a multidimensional feature vector is constructed based on the vibration intensity value it bears, the target velocity value, the target radar cross section value, and the target semantic category; The multidimensional feature vector is input into a preset scoring function, which combines the feature values ​​of each dimension with preset weights to output the multimodal consistency score.

6. The system according to claim 5, characterized in that, The target semantic categories include "person", "vehicle", "animal" and "environmental interference"; the scoring function assigns a higher weight to the "person" semantic category than to other semantic categories.

7. The system according to claim 1, characterized in that, The target active region is a connected subset in a three-dimensional spatial grid whose volume is greater than a preset volume threshold, and whose multimodal consistency scores of all grid cells are higher than the preset score threshold.

8. The system according to claim 1, characterized in that, The system also includes: An environmental baseline adaptive module is used to statistically analyze the distribution of the vibration intensity value and the target point cloud data attributes over multiple historical time periods when no alarm signal is generated, in order to construct an environmental noise baseline model. The environmental baseline adaptive module is also used to dynamically adjust the preset threshold values ​​of vibration intensity and target velocity used by the event kernel identification module according to the environmental noise baseline model.

9. The system according to claim 8, characterized in that, The environmental noise baseline model records the vibration intensity values ​​and statistical characteristics of the target point cloud data attributes under different weather conditions or at different times of the day; the environmental baseline adaptive module selects the corresponding statistical characteristics to adjust the preset threshold based on the current time and weather information.

10. The system according to claim 1, characterized in that, The system also includes: The threat level assessment module is used to comprehensively assess and output the threat level of this intrusion event based on the volume of the target activity area, the average multimodal consistency score of its internal grid cells, and the target semantic category after outputting the perimeter intrusion alarm signal.

Images