An intrusion target rapid identification method for unmanned water station

Abstract

The application relates to the technical field of image processing, and discloses a rapid identification method for an invasion target of an unmanned water station, which comprises the following steps: performing background modeling on a video stream to extract an initial motion candidate frame; assigning an interframe interval number and a ratio threshold value to the candidate frame according to a physical depth map, determining a time window according to the ratio threshold value, calculating a direction consistency ratio of the candidate frame in the time window, and discarding the candidate frame with a direction consistency ratio less than the ratio threshold value to obtain a first reserved candidate frame; mapping the first reserved candidate frame to a physical surface by using a homography matrix to calculate a physical speed, discarding the frame with a speed greater than an upper limit to obtain a second reserved candidate frame, and extracting a physical constraint tuple; outputting a semantic category label by using a convolutional neural network, checking the label with the physical constraint tuple, and triggering an alarm. The application combines physical space rules and image semantic features for double checking, effectively eliminates water surface noise interference, and significantly improves the rapid identification accuracy of the invasion target of the unmanned water station.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for rapid identification of intrusion targets in unmanned water stations. Background Technology

[0002] Unmanned water stations are unattended facilities deployed around natural bodies of water for automatic monitoring of hydrological and water quality data. Intrusion targets primarily refer to unauthorized personnel or vessels entering the protected area of ​​the water station. To ensure the safety of the water station equipment and environment, the security system must possess rapid identification capabilities, meaning it must process surveillance video streams in real time at edge devices to quickly detect and classify unidentified entities for timely response.

[0003] Existing water station security largely relies on on-site video surveillance equipment, combined with basic image motion detection algorithms or general deep learning models for prevention. In conventional applications, the system mainly extracts motion regions by capturing pixel changes between video frames, then directly inputs this image region into a neural network, relying solely on the texture and shape features of the two-dimensional image to output a prediction of the intrusion target category, and sends alarm information to the backend accordingly.

[0004] Conventional techniques of this kind have significant limitations in adapting to complex aquatic environments. Ripples on natural water surfaces, swaying aquatic plants, and changes in light and shadow can easily cause pixel fluctuations, which traditional algorithms readily misidentify as intrusion targets, resulting in a large number of false alarms. More critically, existing recognition models classify based solely on two-dimensional visual semantics, lacking constraints and verification regarding the spatial scale and actual speed of intrusion targets in the real physical world. This single image-level judgment makes it difficult for the system to accurately distinguish between real physical intrusions and environmental noise, leading to a surge in invalid computation. This not only slows down the overall processing speed but also severely restricts the accuracy of monitoring and early warning.

[0005] Therefore, this invention proposes a rapid identification method for intrusion targets in unmanned water stations to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a rapid intrusion target identification method for unmanned water stations. This method solves the problem that existing video surveillance equipment relies solely on two-dimensional image features and lacks verification of real physical spatial parameters, resulting in a high susceptibility to environmental noise interference in complex water environments, leading to numerous false alarms and low identification efficiency.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for rapid identification of intrusion targets in unmanned water stations, comprising the following steps: Background modeling is performed on the video stream to extract initial motion candidate boxes; Based on the physical depth map, the initial motion candidate boxes are assigned a number of frame intervals and a ratio threshold. A time window is determined based on the number of frame intervals. The corresponding continuous time frames within the time window are extracted. The orientation consistency ratio of the initial motion candidate boxes within the time window is calculated. Initial motion candidate boxes with an orientation consistency ratio less than the ratio threshold are discarded to obtain the first retained candidate boxes. The first reserved candidate box is mapped to the physical surface using the homography matrix to calculate the physical velocity. The first reserved candidate box with a physical velocity greater than the velocity upper limit is discarded to obtain the second reserved candidate box. The physical constraint tuple of the second reserved candidate box is extracted. The semantic category label of the second retained candidate box is output using a convolutional neural network. The semantic category label is then verified against the physical constraint tuple, and an alarm is triggered.

[0008] Preferably, the step of extracting initial motion candidate boxes by background modeling of the video stream further includes: A static region mask matrix is ​​established to divide the water surface region and the land region. The homography matrix corresponding to the water surface region and the land region is obtained respectively, resulting in the water surface homography matrix and the land homography matrix. Based on the static region mask matrix, the water surface homography matrix and the land homography matrix are called to calculate the perspective depth and stitch them together to generate the physical depth map.

[0009] The homography matrix is ​​a matrix that characterizes the projection transformation relationship between image pixel coordinates and physical plane coordinates. The water surface homography matrix corresponds to the absolute horizontal plane of the water surface, and the land homography matrix corresponds to the absolute horizontal plane of the land surface, so that pixels in different regions can complete perspective mapping based on the corresponding physical elevation reference.

[0010] Preferably, the step of performing background modeling on the video stream to extract initial motion candidate boxes includes: The decoded video stream frame sequence is obtained and grayscale processing is performed to obtain grayscale image frames; A Gaussian mixture model is used to model the background of the grayscale image frame, separate the foreground pixels, and output a binarized foreground mask image. Perform morphological opening and morphological closing operations on the binarized foreground mask image; The minimum bounding rectangle of independent connected components is calculated using a connected component analysis algorithm. The minimum bounding rectangle with a pixel area smaller than a preset lower limit is removed to obtain the initial motion candidate box, and the center point coordinates of the initial motion candidate box are extracted.

[0011] Preferably, the step of allocating the number of frame intervals and the ratio threshold to the initial motion candidate boxes based on the physical depth map includes: The depth value is obtained from the physical depth map based on the coordinates of the center point. Obtain the pre-defined effective depth determination range, frame interval number extreme range, and ratio threshold extreme range; The number of frame intervals corresponding to the depth value within the extreme range of the number of frame intervals is calculated using a truncated linear mapping and a direct proportional interpolation formula, and the ratio threshold corresponding to the depth value within the extreme range of the ratio threshold is calculated using an inverse proportional interpolation formula.

[0012] Preferably, the step of calculating the orientation consistency ratio of the initial motion candidate boxes within the time window includes: Calculate the average motion vector between each pair of adjacent frames in the time window for the initial motion candidate box; Based on the horizontal and vertical components of all the average motion vectors within the time window, vector displacement synthesis and scalar path accumulation calculations are performed respectively to obtain the vector time integral magnitude and absolute value integral magnitude of the average motion vector. The directional consistency ratio is obtained by dividing the vector time integral magnitude by the sum of the absolute value integral magnitude and the positive constant.

[0013] Preferably, the step of mapping the first reserved candidate box to the physical surface using the homography matrix to calculate the physical velocity includes: Extract the perspective reference points of the first retained candidate box; Determine the region attribute to which the perspective reference point belongs, and match the corresponding homography matrix according to the region attribute; By combining the average motion vector with the perspective reference point pixel coordinates under the current time frame index, the perspective reference point pixel coordinates under the historical reference time frame index are obtained through inverse displacement cumulative calculation. The pixel coordinates of the perspective reference points corresponding to the current time frame and the historical reference time frame are mapped to the physical surface and normalized using the matched homography matrix to obtain the corresponding absolute horizontal physical coordinates, and then the physical velocity is calculated.

[0014] Preferably, the step of discarding the first retained candidate boxes whose physical velocities are greater than the upper speed limit to obtain the second retained candidate boxes, and extracting the physical constraint tuples of the second retained candidate boxes includes: The physical speed is compared with the speed limit, and the first retained candidate box with the physical speed greater than the speed limit is discarded, and the first retained candidate box with the physical speed less than or equal to the speed limit is used as the second retained candidate box; Extract the pixel width and pixel height of the second retained candidate box, as well as the pixel coordinates of the bottom endpoint; The physical width is calculated by mapping the endpoint pixel coordinates to the physical surface using the matched homography matrix; Based on the physical width and the pixel height, and the physical depth height is approximately estimated using the pixel width, the physical projection bottom area is obtained by product. The region attributes, the physical velocity, and the physical projection area are encapsulated to form the physical constraint tuple.

[0015] Preferably, the step of outputting the semantic category label of the second retained candidate box using a convolutional neural network includes: The color pixel cropping region is obtained by using the vertex pixel coordinates of the second candidate box as the cropping boundary to extract the original color image frame; The color pixel cropping region is scaled to a fixed spatial resolution using a bilinear interpolation algorithm and then numerically normalized to generate an image tensor. The image tensor is input into the convolutional neural network for forward propagation calculation to obtain a probability distribution vector. The business category corresponding to the component with the largest value in the probability distribution vector is extracted as the semantic category label.

[0016] Preferably, the step of verifying the semantic category label with the physical constraint tuple further includes: The physical limit constraints corresponding to the semantic category label are obtained by looking up a table in a pre-configured semantic-physical prior database; the physical limit constraints include a reasonable set of regional attributes, a reasonable speed range, and a reasonable bottom area range.

[0017] Preferably, the step of verifying the semantic category label with the physical constraint tuple and triggering an alarm includes: Determine whether the region attribute belongs to the reasonable region attribute set, verify whether the physical speed falls within the reasonable speed range, and determine whether the physical projection bottom area is within the reasonable bottom area range. If all the determination conditions are met, then the target corresponding to the second reserved candidate box is confirmed to be a real physical intrusion entity; The current color original image frame containing the actual physical intrusion entity is captured as a scene capture image. The semantic category label and the physical speed are extracted, and the scene capture image is pushed to the remote management platform and triggers the sound and light alarm device to issue a warning.

[0018] This invention provides a method for rapid identification of intrusion targets in unmanned water stations. It has the following beneficial effects: 1. This invention acquires real-time video streams through video surveillance equipment, employs background modeling and moving target detection technologies to initially locate the activity area, and filters candidate bounding boxes by combining physical depth maps and orientation consistency ratios. This design effectively eliminates interference from natural environmental factors such as water ripples around unmanned water stations, accurately locates suspected intrusion targets, reduces subsequent unnecessary computational overhead, thus laying a data foundation for rapid identification of the entire system and improving the accuracy of initial detection.

[0019] 2. This invention maps candidate boxes to physical surfaces using a homography matrix to extract physical constraint tuples such as physical velocity, and then uses a lightweight convolutional neural network model to output category labels, performing logical verification between the two. This joint verification method of visual semantics and physical laws can accurately identify real threats within the unmanned water station's defense area, effectively reducing false alarms for non-intrusive targets. Simultaneously, the application of a lightweight network ensures efficient forward propagation for rapid identification.

[0020] 3. This invention, upon completing the semantic and physical parameter matching verification of the image, immediately triggers the local audible and visual alarm device deployed on-site, and simultaneously pushes the alarm information and captured images to the remote management platform. This linkage mechanism enables immediate on-site warning and deportation upon confirmation of the intrusion target, and reports the system's rapid identification and judgment results to management personnel, achieving a closed-loop monitoring system from visual perception to danger alarm, effectively ensuring the safety of unmanned water stations. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for rapid identification of intrusion targets in an unmanned water station according to the present invention; Figure 2 This is an architecture diagram of an intrusion target rapid identification system for an unmanned water station according to the present invention; Figure 3 This is a flowchart illustrating the semantic category label and physical constraint tuple verification process of the present invention. Figure 4 This is a line graph showing the cumulative change of the vector magnitude of the present invention with the number of frames in the time window; Figure 5 This is a cross-plot of the distribution mapping of the target physical characteristic parameters of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 This invention provides a method for rapid identification of intrusion targets in unmanned water stations, comprising the following steps: S1, perform background modeling on the video stream to extract initial motion candidate boxes; S2. Based on the physical depth map, assign the number of frame intervals and the ratio threshold to the initial motion candidate boxes. Determine the time window based on the number of frame intervals. Extract the corresponding continuous time frames within the time window. Calculate the orientation consistency ratio of the initial motion candidate boxes within the time window. Discard the initial motion candidate boxes whose orientation consistency ratio is less than the ratio threshold to obtain the first retained candidate boxes. S3, use the homography matrix to map the first retained candidate box to the physical surface to calculate the physical velocity, discard the first retained candidate box whose physical velocity is greater than the velocity upper limit to obtain the second retained candidate box, and extract the physical constraint tuple of the second retained candidate box; S4. Use a convolutional neural network to output the semantic category label of the second retained candidate box, verify the semantic category label with the physical constraint tuple and trigger an alarm.

[0024] See attached document Figure 2 This invention provides a rapid intrusion target identification system for unmanned water stations. At the hardware deployment level, the identification system primarily relies on edge computing nodes deployed within the local computer room of the unmanned water station. The system front-end includes video surveillance equipment deployed inside and around the water station, used to continuously acquire real-time video streams with a wide field of view and input them to the edge computing nodes. The logical architecture of the identification system includes: video surveillance equipment, a video preprocessing module, a temporal optical flow filtering module, a perspective motion filtering module, and a cross-domain verification and alarm module.

[0025] The rapid intrusion target identification method and identification system for unmanned water stations of the present invention belong to the same inventive concept. Each logical module in the identification system is configured in an edge computing node to execute the corresponding steps in the identification method. Specifically, the video preprocessing module in the identification system is used to perform step S1 to perform background modeling of the video stream and extract initial motion candidate boxes; the temporal optical flow filtering module is used to perform step S2 to allocate frame intervals and ratio thresholds to the initial motion candidate boxes based on the physical depth map, determine the time window based on the frame intervals, extract the corresponding continuous time frames within the time window, calculate the directional consistency ratio of the initial motion candidate boxes within the time window, discard initial motion candidate boxes with directional consistency ratios less than the ratio threshold to obtain the first retained candidate boxes; the perspective motion filtering module is used to perform step S3 to map the first retained candidate boxes to physical surfaces using a homography matrix to calculate physical velocities, discard the first retained candidate boxes with physical velocities greater than the speed upper limit to obtain the second retained candidate boxes, and extract the physical constraint tuples of the second retained candidate boxes; the cross-domain verification and alarm module is used to perform step S4 to output the semantic category labels of the second retained candidate boxes using a convolutional neural network, verify the semantic category labels against the physical constraint tuples, and trigger an alarm. The recognition system relies on the sequential interaction of each module to achieve a computational closed loop from video preprocessing and physical prior filtering to cross-domain verification and recognition of intrusion targets.

[0026] Before extracting initial motion candidate boxes by background modeling of the video stream, a static region mask matrix needs to be established to divide the water surface region and the land region. The homography matrices corresponding to the water surface region and the land region are then obtained, resulting in water surface homography matrices and land homography matrices. Based on the static region mask matrix, the perspective depth is calculated using the water surface homography matrix and the land homography matrix, and then stitched together to generate a physical depth map. The water surface region and the land region refer to the image portions corresponding to the water surface and the land on the shore in the unmanned water station's monitoring field of view, respectively. Since there is a significant elevation difference between the two in the real three-dimensional physical world, using the same reference plane for distance measurement would lead to severe computational distortion. In this embodiment, during the initialization phase, the system generates a two-dimensional array with the same resolution as the monitoring image by manually calibrating or calling a semantic segmentation algorithm to statically analyze the scene; this array is the static region mask matrix. In this static region mask matrix, pixels with a value of 0 are identified as belonging to the water surface region, and pixels with a value of 1 are identified as belonging to the land region. For these two physical planes with different elevations, the system obtains the water surface homography matrix and the land homography matrix respectively. The water surface homography matrix and the land homography matrix are 3×3 geometric perspective transformation matrices used to map and transform the pixel coordinate system in a two-dimensional surveillance image to the absolute horizontal planes of water and land in the real three-dimensional world, respectively. These two matrices, by incorporating camera intrinsic and extrinsic parameters, can compensate for perspective distortion caused by objects appearing larger when closer and smaller when farther away, restoring the absolute distance and relative position of pixels in actual physical space. Those skilled in the art can obtain these two matrices using pre-set calibration references and the Zhang Zhengyou calibration method; this is a well-known technique. Subsequently, based on the static region mask matrix, the water surface homography matrix and the land homography matrix are used to calculate the perspective depth and stitch them together to generate a physical depth map. The physical depth map is a floating-point matrix recording the actual physical distance from each pixel to the camera's optical center. The generation process is as follows: The static region mask matrix is ​​traversed pixel by pixel. When a pixel value indicates a water surface region, the current pixel coordinates are extracted and a coordinate system transformation is performed using the water surface homography matrix. The true 3D spatial coordinates of the point under the water surface elevation are calculated, and then the straight-line distance between the point and the camera's optical center is obtained as the perspective depth. Similarly, when the indicated region is land, the land homography matrix is ​​called for the same calculation. This 3D coordinate projection and distance calculation process based on the homography matrix is ​​a conventional existing technique in perspective geometry. Finally, the system arranges all the calculated perspective depths according to their original 2D spatial positions and stitches them together to obtain a physical depth map for subsequent steps.

[0027] See attached document Figure 1 and attached Figure 2 The specific execution logic for step S1 includes the following sub-steps: S101, Obtain the video stream frame sequence decoded by the edge computing node. In this embodiment, the edge computing node reads the real-time video stream collected by video surveillance equipment deployed inside and around the unmanned water station through a video access protocol, generating a video stream frame sequence. As a preferred method, the video surveillance equipment can be a high-definition network bullet camera or dome camera that supports network video transmission protocols; the real-time video stream is an H.264 or H.265 format video stream containing continuous timestamps and image frame data. The edge computing node decodes the video stream frame by frame, extracting continuous color original image frames. This video stream decoding and extraction process is a well-known technology in the field. For example, those skilled in the art can use the video reading interface (VideoCapture class) in an open-source computer vision library (such as OpenCV) or the FFmpeg multimedia processing framework to parse and restore the compressed video stream into a two-dimensional pixel array with RGB three-channel characteristics in chronological order. This two-dimensional pixel array is the color original image frame. To reduce the subsequent matrix operation and unify the data format, each color original image frame is grayscaled to obtain grayscale image frames I arranged in chronological order. t (u,v), where t is the index of the current time frame, and (u,v) represents the pixel coordinates of a pixel in the original color image frame; simultaneously, the corresponding original color image frame is cached in the memory of the edge computing node for subsequent feature extraction. This grayscale processing is a well-known technique in the field of computer vision. Those skilled in the art can, for example, use a weighted average method to convert the RGB three-channel pixel values ​​into single-channel grayscale values, thereby eliminating redundant color information.

[0028] S102, using a Gaussian mixture model to process grayscale image frame I t Pixel-level background modeling is performed for (u,v) to separate foreground pixels. Faced with the complex environment of unmanned water stations, a single background model often struggles to adapt to drastic changes in water ripples and lighting. By establishing multiple Gaussian mixture models with Gaussian distributions for each pixel, the statistical regularities of the dynamic background can be effectively characterized. Specifically, for any pixel (u,v) in a grayscale image frame, the pixel value X at the current time frame index t... t Its observation probability P(X) t The calculation formula for ) is as follows: ; Where K is the preset number of Gaussian distributions, and in this embodiment, the value of K is set to be between 3 and 5; ω i,t μ represents the weight of the i-th Gaussian distribution at the current time frame index t. i,t σ is the scalar mean of the i-th Gaussian distribution at the current time frame index t (degenerates to a one-dimensional scalar here due to the previous grayscale conversion); i,t 2Let η be the variance of the i-th Gaussian distribution at the current time frame index t; η() is the Gaussian probability density function.

[0029] Observation probability P(X) t ) is used to represent the current pixel value X t The probability generated by the overall background model serves as a macroscopic basis for measuring whether a pixel conforms to historical background statistical patterns.

[0030] In the specific foreground determination process, the above formula indicates that the Gaussian mixture model is a weighted sum of K individual Gaussian distributions. The Gaussian mixture model first sorts these K Gaussian distributions according to their weights, considering the distribution with the higher weight as the background distribution representing the current reliable background state. Subsequently, the Gaussian mixture model calculates the pixel value X of each pixel... t The pixel is matched against these Gaussian distributions that represent background states one by one. When the pixel value deviates from the mean of all these background distributions by more than a set threshold (e.g., within 2.5 times the standard deviation), the pixel is determined not to belong to any known background state and is thus marked as a foreground pixel.

[0031] For the parameter update mechanism of the Gaussian mixture model, those skilled in the art can make conventional settings according to the lighting conditions of the actual scene. The relevant mathematical derivation and parameter optimization are well-known techniques in this field and will not be elaborated here. After the full-image pixel-level determination of the grayscale image frame, the system outputs a binarized foreground mask image containing foreground and background pixels.

[0032] S103, morphological filtering is performed on the binarized foreground mask image. Due to the presence of swaying aquatic plants and fine reflections on the water surface at the water station site, the previously separated foreground pixels typically contain a large number of discrete noise points. At this stage, the system performs morphological opening operations on the binarized foreground mask image to eliminate isolated noise pixel blocks, and performs morphological closing operations to fill pixel holes inside the actual target, making the edges of the foreground region smoother. Morphological opening and closing operations are well-known techniques in this field.

[0033] For example, when the reflection of water ripples appears as scattered white dots with an area of ​​only a few pixels in a binarized foreground mask image, it can be eroded away by setting the opening operation of the structuring element with a specific size (e.g., using a 3×3 or 5×5 rectangle); while when the reflection of a real ship causes black holes with internal fractures in the image, the closing operation can use the dilation property to fill these holes and restore the complete outline of the target.

[0034] S104, Initial motion candidate boxes are extracted using a connected component analysis algorithm. In the binarized foreground mask image after morphological filtering, the system searches for sets of interconnected foreground pixels and marks them as independent connected components. For each independent connected component, the minimum bounding rectangle containing all pixels of that independent connected component is calculated. To filter out interference items with excessively small areas, the pixel area of ​​this minimum bounding rectangle is calculated, and minimum bounding rectangles with pixel areas smaller than a preset lower limit are discarded. This preset lower limit can be empirically set based on the field of view of the monitoring screen and the expected minimum physical size of the target, for example, set to 100 pixels. After removing invalid interference, the remaining minimum bounding rectangles are defined as initial motion candidate boxes. The vertex pixel coordinates and center point coordinates (u) of each initial motion candidate box are further recorded. c ,v c If the calculated center point coordinates are floating-point numbers, they are rounded to the nearest integer to meet the integer indexing requirements of the matrix. The center point coordinates (u...) c ,v c After extraction, the initial motion candidate boxes are cached in the memory of the edge computing nodes and used as spatial location indexes in subsequent processing to establish a mapping relationship with the physical depth map. During this process, if no initial motion candidate boxes satisfying the area condition are extracted from the current color raw image frame, the system directly discards the current color raw image frame and waits for the next frame of video stream input, thus avoiding unnecessary occupation of subsequent computing resources. All initial motion candidate boxes extracted within the current color raw image frame constitute the candidate box set for that time node, serving as the underlying data input for subsequent spatiotemporal dual filtering.

[0035] See attached document Figure 1 and attached Figure 2 The specific execution logic for step S2 includes the following sub-steps: S201, extract the center point coordinates of the initial motion candidate boxes, and obtain the depth value in the physical depth map based on the center point coordinates. For each initial motion candidate box extracted in step S1, cache its center point coordinates (u... c ,v c Using this as a lookup index, the corresponding value is directly extracted from the physical depth map. This value represents the depth of the initial motion candidate box in the current physical space.

[0036] S202, based on the physical depth map, assigns frame intervals and ratio thresholds to the initial motion candidate boxes. The logic is to assign larger frame intervals and smaller ratio thresholds to depth values. Since distant regions produce smaller pixel displacements on the image plane, longer observation times are needed to accumulate significant motion features, and targets in distant regions are significantly affected by pixel quantization noise, the false alarm threshold needs to be appropriately relaxed. Conversely, targets in close-range regions have large pixel areas and move rapidly, requiring extremely fast system responses. As a preferred approach, the system pre-sets the effective judgment range [D] for depth values. min D max ], and the corresponding extreme range of frame interval number [T min ,T max (e.g., set to an interval of 3 to 15 adjacent frames) and the extreme range of the ratio threshold [R] min ,R max (For example, set to 0.2 to 0.8). Where, D min With D max These are the preset minimum and maximum effective depth limits, respectively. Their value range can be set according to the effective monitoring line distance of the on-site camera, for example, D. min The value is 5 meters, D max The value is 150 meters; T min With T max These are the preset minimum and maximum frame interval number extreme values, respectively; R min With R max These are the preset minimum and maximum ratio threshold extremes, respectively. When the depth value D of an initial motion candidate box is obtained, the corresponding frame interval N is calculated using a truncated linear mapping function. T With ratio threshold R th Specifically, if D <D min Let D=D min If D>D max Then let D=D max Then, the direct proportional interpolation formula N was used. T =floor(T min +(T max -T min )×(DD min ) / (D max -D min The number of frame intervals is calculated and allocated. floor() is a floor function that rounds down, and the inverse proportional interpolation formula R is used. th =R max -(R max -R min )×(DD min ) / (Dmax -D min The ratio threshold is calculated and assigned. Through the above numerical mapping relationship, each initial motion candidate box is assigned temporal analysis parameters that match its spatial position.

[0037] S203, calculate the average motion vector of the initial motion candidate boxes within the time window. This is based on the allocated number of frame intervals N. T Using the current time frame index t as the ending frame index, and tN T Starting with the frame index tN, extract the frame indices tN sequentially from the cached grayscale image frame sequence. T ,tN T +1, ..., t-1, t, and the continuous N T +1 grayscale images are used, arranged chronologically to form a time window for analyzing the initial motion candidate box. Within this time window, the system calculates the dense optical flow field for the two-dimensional pixel region corresponding to the current initial motion candidate box between each pair of adjacent frames. Those skilled in the art can use the Farneback polynomial expansion algorithm to calculate the horizontal and vertical optical flow velocity components of each pixel within the two-dimensional pixel region; this is a well-known technique. Based on this, the arithmetic mean of the optical flow velocity components of all foreground pixels within the initial motion candidate box region is calculated. Since the time window includes N... T +1 frame of grayscale image, with N frames between adjacent frames. T Therefore, N can be calculated. T A dense optical flow field, and further obtained N T Let the i-th average motion vector be V. x (i), the vertical component is V y (i), where i is the traversal index of the average motion vector between adjacent frames, and its value ranges from 1 to i to N. T .

[0038] S204 calculates the vector time integral magnitude and the absolute value integral magnitude of the average motion vector. To quantify the spatial consistency characteristics of the motion trajectory, the system first defines the vector time integral magnitude M. int It represents the length of the target's actual net spatial displacement vector within the current time window; at the same time, it defines the absolute value integral magnitude M. abs This represents the total scalar distance traveled by the target's trajectory within the same time window. Regarding the previously obtained N... T The average motion vector is used to calculate the vector time integral magnitude M using the following formula. int : ; At the same time, the absolute value integral modulus M is calculated using the following formula. abs : .

[0039] S205, the vector time integral magnitude is divided by the absolute value integral magnitude to obtain the orientation consistency ratio. Initial motion candidate boxes with orientation consistency ratios less than a ratio threshold are discarded to obtain the first retained candidate boxes. Based on the two obtained magnitude values, the system calculates the ratio M. int / (M abs +ε) to obtain a directional consistency ratio between 0 and 1, where ε is a very small positive constant (e.g., 10) introduced to prevent system crashes caused by a denominator of zero. -5 When the target in the video frame is high-frequency ripples on the water surface or water plants swaying in the wind, it generates high-frequency reciprocating oscillations within the current time window. The positive and negative motion vectors cancel each other out during accumulation, resulting in M. int Approaching zero, while M abs However, because the cumulative displacement distance is not zero, the directional consistency ratio is significantly low. The system compares the calculated directional consistency ratio with the previously assigned ratio threshold R. th The system compares the directional consistency ratios. If the ratio is less than a threshold, the target is identified as a pseudo-target caused by high-frequency oscillations in place, a non-physical intrusion, and the initial motion candidate box is discarded. Conversely, if the directional consistency ratio is greater than or equal to the threshold, it indicates that the target has a continuous and unidirectional physical displacement trajectory. The system retains this target and defines it as the first reserved candidate box, used as data input for subsequent step S3. This computational model based on local spatiotemporal constraints filters out most of the dynamic environmental interference in the water station scenario from the perspective of the physical laws of temporal evolution without relying on high-performance feature extraction.

[0040] See attached document Figure 1 and attached Figure 2 The specific execution logic for step S3 includes the following sub-steps: S301, Extract the perspective reference point of the first retained candidate box. After obtaining the first retained candidate box after spatiotemporal dual filtering, the system needs to select the key pixel point representing the target position for spatial mapping. Since the contact surface between a moving target (e.g., a moving ship or a walking person) and its physical plane (water or land) is usually reflected in the bottom edge of its pixel bounding box, selecting this location can minimize the projection ranging error caused by the target's own height. In this embodiment, the center pixel point of the line segment where the bottom edge of the first retained candidate box is located is extracted as the perspective reference point, and the pixel coordinates (u) of this perspective reference point in the current color original image frame are recorded. b ,v b ).

[0041] S302, determine the region attribute to which the perspective reference point belongs, and match the corresponding homography matrix based on this region attribute. Set the pixel coordinates (u) of the perspective reference point... b ,v b The system uses a 0 as an index to query the static region mask matrix established during the initialization phase. If the retrieved matrix value is 0, the region attribute of the perspective reference point is determined to be a water surface region, and the system adaptively routes and calls the water surface homography matrix; if the matrix value is 1, the region attribute is determined to be a land region, and the system adaptively routes and calls the land homography matrix; then, the currently determined region attribute is denoted as Attr. This spatial location-based adaptive routing mechanism ensures that cross-domain targets can be processed based on the correct physical elevation reference surface. For ease of mathematical expression, the 3×3 homography matrix actually called based on the region attribute is uniformly denoted as H.

[0042] S303, the first reserved candidate box is mapped to the physical plane using the homography matrix to calculate the physical velocity. The current time frame index is set to t, and the first reserved candidate box calculated in step S203 is used in conjunction with the number of frame intervals N. T N within the current time window T A plurality of average motion vectors are used to perform inverse displacement and cumulative subtraction calculations on the perspective reference point pixel coordinates under the current time frame index t, thereby obtaining the perspective reference point pixel coordinates of the target represented by the first retained candidate box in the historical reference time frame index. As a preferred method, the historical reference time frame is the one based on the number of frame intervals N in the aforementioned step S2. T The starting frame of the current time window is determined; the index of this historical reference time frame is set to t. ref It is equal to the current time frame index t minus the number of frame intervals N. T The specific process of reverse displacement cumulative calculation is as follows: Set the pixel coordinates of the perspective reference point under the current time frame index t as (u b,t ,v b,t The horizontal component of the i-th average motion vector within the current time window obtained in step S203 above is V. x (i) The vertical component is V y (i) then the target is at the historical reference time frame index t ref perspective reference point pixel coordinates The frame-by-frame displacement reduction can be calculated using the following formula: ; ; This completes the derivation of the spatial location of the same target across frames.

[0043] Furthermore, the absolute time interval between the current time frame and the historical reference time frame is defined as Δt, the value of which is equal to the number of frame intervals N. T Multiply by the camera's preset single-frame capture duration (e.g., 0.04 seconds when the frame rate is 25fps). Construct a three-dimensional homogeneous coordinate column vector P from the perspective reference point pixel coordinates at the current time frame index t. t =[u b,t ,v b,t ,1] T The mapping to the absolute physical surface is performed using matrix multiplication: Q t =HP t ; In the formula, Q t Let Q be the mapped 3D homogeneous coordinate column vector. t =[x t ',y t ',z t '] T To obtain the absolute horizontal physical coordinates (X, Y) of this perspective reference point on a two-dimensional real physical plane. t ,Y t Normalization is performed using perspective division: X t =x t ' / z t '; Y t =y t ' / z t '; Similarly, the same homography matrix H is used to calculate the perspective reference point at the historical reference time frame index t. ref The absolute horizontal physical coordinates (X) below ref ,Y ref ).

[0044] Based on the two absolute horizontal physical coordinates obtained above, the physical velocity V of the target represented by the first reserved candidate box is calculated. phy : ; S304: Perform a physical velocity upper limit check, discarding the first candidate bounding box whose physical velocity exceeds the upper limit and obtaining the second candidate bounding box. In the actual monitoring environment of an unmanned water station, birds or insects flying close to the camera often exhibit pixel displacement. Since they are not actually in contact with the water or land surface, when the system forces their perspective reference point to be mapped to a distant physical surface based on the homography matrix of the water or land surface, an abnormally distorted physical velocity will be calculated. A global physical velocity upper limit V is pre-set based on the motion limits of motorized vessels and non-motorized personnel that may appear at the water station.limit For example, V limit The value can be set to 25 meters per second. The physical velocity V, calculated in real time, will be used. phy With this speed limit V limit Perform a comparison, and when V is determined phy >V limit If the target is confirmed to be a false target caused by perspective distortion, it is discarded directly; only the first candidate box with a physical velocity less than or equal to the upper limit of that velocity is retained, and it is used as the second candidate box after preliminary verification by physics.

[0045] S305, Extract the physical constraint tuples of the second retained candidate boxes. After completing the velocity filtering, in order to provide multi-dimensional physical reference data for subsequent steps, the system simultaneously calculates other key physical quantities for the retained second candidate boxes. Specifically, it calculates the physical projected base area S of the target. phy In this embodiment, the system extracts the pixel width w of the second retained candidate box. p With pixel height h p Extract the left endpoint pixel coordinates (u) of the bottom edge of the second retained candidate box. l ,v b ) and the right endpoint pixel coordinates (u r ,v b Each of these is mapped to a physical surface using the homography matrix H obtained from the previous call, thus obtaining the corresponding absolute physical coordinates (X). l ,Y l ) and (X r ,Y r Based on this, the true physical width of the target on the physical plane is calculated. Based on the principle of perspective similarity, the physical depth height H of the target in three-dimensional space is approximately estimated. phy =W phy ×(h p / w p Finally, the physical projection base area S is obtained by quadrating. phy =W phy ×H phy The system then uses the previously queried area attribute Attr and the currently calculated physical velocity V. phy and the physical projection base area S phy The data is encapsulated into a dataset with a fixed structure, forming a physical constraint tuple. This physical constraint tuple is uniquely bound to the image data of the second retained candidate box in memory to support subsequent deep conflict verification.

[0046] See attached document Figure 1 To be continued Figure 3 The specific execution logic for step S4 includes the following sub-steps: S401: Obtain the image data corresponding to the second candidate bounding box and perform size and numerical preprocessing to generate an image tensor adapted to the input of the convolutional neural network. Extract the vertex pixel coordinates of the second candidate bounding box. Since the aforementioned steps (S1 to S3) are mainly based on single-channel grayscale images and binarized masks for feature calculation, to ensure the integrity of the image color features input to the convolutional neural network, the vertex pixel coordinates of the second candidate bounding box are used as the cropping boundary. Local cropping is directly performed on the current color original image frame synchronously cached in step S101 to obtain the color pixel cropping region containing the target. Since the size of the second candidate bounding box of different targets varies in the image, the color pixel cropping region is uniformly scaled to a fixed spatial resolution using a bilinear interpolation algorithm, for example, a preset 224×224 pixels. Subtract the pre-statistical RGB channel mean and divide by the channel standard deviation for each pixel of the scaled image to complete the numerical normalization of the pixel distribution. After the above preprocessing, the two-dimensional image is converted into a three-channel floating-point tensor data with dimensions of 3×224×224. This three-channel floating-point tensor data is the image tensor, which is used as the standardized input for subsequent network inference.

[0047] S402, the image tensor is processed using a convolutional neural network to output the semantic category label of the second retained candidate box. The image tensor is input into the convolutional neural network pre-loaded in the memory of the edge computing node for forward propagation computation. In this embodiment, the convolutional neural network specifically adopts the ResNet-18 layer architecture, and its internal layer structure and data flow specifically include: the input layer receives an image tensor with a dimension of 3×224×224; it first passes through an initial convolutional layer with a kernel size of 7×7, a stride of 2, and 64 output channels, followed by a 3×3 max pooling layer to downsample the feature map to 64×56×56 to extract shallow texture features; then, four residual modules are sequentially cascaded, and the number of output channels of these four combinations are configured as 64, 128, 256, and 512, respectively. Each residual module combination consists of two basic residual blocks, each containing two consecutive 3×3 convolutional layers. A skip connection is used to add the corresponding element-wise input feature tensor of the basic residual block to the output tensor of its second convolutional layer. This structure effectively alleviates the gradient vanishing problem during deep network training and extracts deep abstract semantic features. After four residual module combinations, the output is a 512×7×7 feature map. A global average pooling layer then compresses the spatial feature map into a 512-dimensional one-dimensional feature vector. Finally, a fully connected layer is connected, and a Softmax activation function is applied for probability mapping. The output of the convolutional neural network is a set of probability distribution vectors corresponding to preset business categories. These categories specifically cover common monitored entities at water stations, including motorized vessels, non-motorized vessels, personnel, floating objects, and birds. The business category corresponding to the largest component in this probability distribution vector is extracted and defined as the semantic category label of the second retained candidate box.

[0048] To enable convolutional neural networks (CNNs) to accurately classify data, a preliminary model training step is required. In the water station monitoring scenario of this invention, historical monitoring images containing the aforementioned monitoring entities are pre-collected and cropped. The true business category of each monitoring entity in each historical image is manually labeled, and this label is converted into a true category label vector using one-hot encoding, which is then packaged as a training sample set. During the training phase, the image tensors in the training sample set are input into the pre-constructed CNN to calculate the predicted probability distribution vector. Then, the cross-entropy loss function is used to quantify the error between the predicted probability distribution vector and the true category label vector. Based on this error, the parameter gradients of each hidden layer of the CNN are calculated using the backpropagation algorithm. The Adam optimizer is then called to update the weight parameters within the CNN according to a set learning rate. This forward propagation, backpropagation, and parameter update process is repeated iteratively until the cross-entropy loss function value decreases below the convergence threshold. At this point, the final weight parameters of the CNN are saved for real-time monitoring and inference calls on edge computing nodes. The above-described backpropagation-based neural network training process is well-known in the art and will not be elaborated further here.

[0049] S403, Obtain the physical limit constraints corresponding to the semantic category label. In this embodiment, a semantic-physical prior database is pre-configured in the storage medium of the edge computing node. This semantic-physical prior database is essentially a mapping table that records the objective parameter extreme values ​​that various semantic entities should follow in the real three-dimensional physical world. Based on the output semantic category label, the bound physical limit constraints are obtained by looking up the table. The physical limit constraints specifically include: a reasonable region attribute set C. attr Reasonable speed range [V] min V max ] and reasonable bottom area range [S min ,S max ]. Among them, C attr This is a set containing mask identifier values, representing the area (such as water or land) where such targets might exist under normal physical conditions. It should be clearly noted that the region attribute Attr calculated in the preceding steps is a single attribute state value extracted for the actual location of a specific target in the image, while the reasonable region attribute set C... attr It is the set of all legal attribute state values ​​allowed to exist in the current semantic category under macroscopic physical laws; the two belong to the logical verification relationship between element and set; V min With V max These represent the minimum and maximum physical velocity extrema of this type of target on a real physical surface, respectively; S min With S maxThese represent the minimum and maximum physical projection base areas of this type of target in three-dimensional space, respectively. For example, when the output semantic category label is "person," real humans typically move along the shore at a relatively slow speed, and their corresponding reasonable region attribute set C... attr The system is configured to include only the identifier value 1, representing land areas; the reasonable speed range is set to 0 to 10 meters per second; and the reasonable bottom area range is set to 0.05 to 1.0 square meters. When the semantic category label is "motorized vessel," its actual operating area is in water and it is large in size, so the reasonable area attribute set C is... attr The setting is to include only the identifier value 0 representing the water surface area, the reasonable speed range is set to 2 to 20 meters per second, and the reasonable bottom area range is set to 2.0 to 50.0 square meters.

[0050] S404, validate the semantic category label against the physical constraint tuple. Retrieve the physical constraint tuple extracted in the previous steps that is uniquely bound to the second retained candidate box. This physical constraint tuple contains the region attribute Attr and the physical velocity V. phy and the physical projection base area S phy These are all physical values ​​calculated in real-time by the edge computing nodes in step S3 above, based on the pixel coordinates of the second retained candidate box in the image and the perspective mapping of the homography matrix to the real physical surface. These real-time physical values ​​are then logically compared with the physical limit constraints obtained in the previous step. The specific verification rule is: determine whether the region attribute Attr belongs to a reasonable region attribute set C. attr Verify physical velocity V phy Does it fall within a reasonable speed range? [V] min V max Within [the area], determine the physical projection base area S. phy Is it within a reasonable range of base area [S]? min ,S max If any of the above judgments is negative, it indicates a direct conflict between the semantic category identified by the image and its three-dimensional physical properties. In real-world environments, this conflict is often caused by irregular reflections from strong water reflections or raindrops adhering to surveillance cameras. These visual distractions happen to resemble the outline of the target of interest in a two-dimensional image, leading the convolutional neural network to output incorrect classification results. However, their elevation, speed, and size in physical space differ significantly from the actual physical kinematic characteristics. When such cross-domain feature conflicts occur, the target corresponding to the second retained candidate box is confirmed as a pseudo-target that is not a real intrusion, and it is directly discarded, terminating subsequent processing.

[0051] S405, trigger closed-loop alarm operation. When all parameters in the physical constraint tuple fully satisfy the physical limit constraints corresponding to the semantic category label, the target achieves logical consistency at the image semantic level and the spatiotemporal physical law level, confirming that the target corresponding to the second reserved candidate box is a real physical intrusion entity. Subsequently, the edge computing node extracts the physical velocity V of the physical intrusion entity from the corresponding physical constraint tuple. phy The current color original image frame containing the physical intrusion entity is captured as the scene snapshot image. Semantic category labels and physical velocity V are extracted. phy The system also captures images on-site and packages them into an alarm data stream in a specific format (such as an alarm message in JSON format). Through the communication network interface built into the edge computing node, the alarm data stream is synchronously pushed to a remote management platform (such as the security monitoring room terminal platform of a reservoir management station). This triggers local audible and visual alarm devices deployed at the unmanned water station (such as a combination of a high-decibel explosion-proof loudspeaker and LED flashing warning lights) to issue a warning, thereby driving away intruders and notifying remote management personnel as soon as possible, ensuring the safety of the unmanned water station, and thus completing the entire monitoring process from visual perception to hazard alarm.

[0052] To further illustrate the implementation process and technical effects of the present invention, the following description is provided in conjunction with specific application scenarios. The specific numerical values ​​and scenario parameters described below are only used to explain the implementation principle of the present invention and do not limit the scope of protection of the present invention.

[0053] Using a monitoring scenario of an unmanned water station in a remote reservoir as an example, this paper explains the specific execution process of a method for rapid identification of intrusion targets at an unmanned water station.

[0054] System Initialization Phase: In this embodiment, the video surveillance equipment deployed inside and around the unmanned water station uses high-definition network bullet cameras. During the initialization phase, the system statically analyzes the scene and generates a static region mask matrix in the form of a two-dimensional array. Pixels with a value of 0 are identified as belonging to the water surface area, and pixels with a value of 1 are identified as belonging to the land area. The system obtains the homography matrix of the water surface and the homography matrix of the land area, and calculates the perspective depth based on the corresponding homography matrix according to the static region mask matrix. The resulting matrix is ​​then stitched together to generate a physical depth map, which is cached as a floating-point matrix in the edge computing nodes.

[0055] Step S1: After the edge computing node acquires the video stream frame sequence, it extracts the original color image frames through decoding and simultaneously performs grayscale processing to obtain grayscale image frame I. t (u,v). For targets appearing in the image, the system uses a Gaussian mixture model to separate foreground pixels. For any pixel (u,v) at the current time frame index t, the grayscale value X... t Its observation probability P(X) tThe calculation formula for ) is as follows: ; After matching and comparison, the system outputs a binarized foreground mask image. The system then performs morphological opening and closing operations on the binarized foreground mask image and uses a connected component analysis algorithm to calculate the minimum bounding rectangle of each independent connected component. A lower limit of 100 pixels is set for the area. After removing minimum bounding rectangles with pixel areas smaller than this lower limit, initial motion candidate boxes for two targets are extracted from this frame: Target A: The person actually walking on the shore; record the calculated coordinates of their center point and round them to obtain integer index coordinates (u). c1 ,v c1 ); Target B: Located on the water surface, experiencing high dynamic disturbance caused by swaying aquatic plants. Record the calculated coordinates of its center point and round them to obtain integer index coordinates (u). c2 ,v c2 ).

[0056] Step S2: Based on the extracted center point coordinates (u c1 ,v c1 ) and (u c2 ,v c2 The system extracts the corresponding depth value D from the physical depth map. The depth value obtained for target A is 15 meters; the depth value obtained for target B is 60 meters. The defined extreme value range is: D. min =5,D max =150; T min =3,T max =15; R min =0.2, R max =0.8. The system uses truncated linear mapping and interpolation formulas for dynamic parameter allocation: where N T The interval between adjacent frames is represented by N. T A defined time window contains N T +1 consecutive time frames. For target A, substituting into the formula, we can calculate the number of adjacent frame intervals N allocated to it. T =3, corresponding to a time window of 4 consecutive frames, ratio threshold R th ≈0.76; For target B, the number N of adjacent frame intervals allocated to it is calculated. T =7, corresponding to a time window of 8 consecutive frames, ratio threshold R th ≈0.57. Within their respective consecutive time frames, the system calculates the horizontal component V of the average motion vector frame by frame. x (i) with vertical component V y (i), and calculate the vector time integral magnitude M. int With absolute value integral modulus Mabs The system assigns N to each of target A and target B. T R th And the average motion vector is bound to the corresponding initial motion candidate box in a cache. For target B (swaying water plants), since its forward and reverse motion vectors cancel each other out during time accumulation, the calculated M... int Approaching zero. When the system accumulates to the 8th frame of its dedicated time window, the calculated directional consistency ratio M is... int / (M abs When the ratio (+ε) is close to 0, which is significantly less than the assigned ratio threshold of 0.57, the system directly discards the initial motion candidate box for target B. For target A (walking person), its ratio is close to 1.0 in the 4th frame, which is greater than the threshold of 0.76, so the system retains it and obtains the first retained candidate box.

[0057] As attached Figure 4 As shown in the figure, this graph visually reflects the difference in integral modulus between target A and target B. Figure 4 The solid black line with circular data points represents the vector time integral magnitude of target A, which increases steadily and linearly with the number of frames. The solid black line with asterisk data points represents the absolute integral magnitude of target A. The two solid black lines almost overlap, indicating that the movement direction of target A is extremely consistent. Conversely... Figure 4 The dark gray dashed line represents target B, whose vector temporal integral magnitude (square data points) oscillates between 0 and 2 pixels and does not accumulate over time; while the absolute value integral magnitude (diamond data points) continuously increases with the number of frames. The huge numerical difference between the two causes the directional consistency ratio of target B to be far below the ratio threshold assigned by the system, and thus it is discarded.

[0058] Step S3: The system extracts the center pixel of the bottom edge line segment of the first candidate bounding box of target A as the perspective reference point, and obtains its pixel coordinates (u b ,v b ). will (u b ,v b The system uses the static region mask matrix as an index to query the region. A matrix value of 1 indicates that the region attribute Attr is land (value 1), and the system calls the land homography matrix H. Frame-by-frame displacement reduction is calculated using the average motion vector, and coordinate mapping is performed using perspective division. The physical velocity V is then calculated. phy =1.6 m / s. Because V phy Less than the set speed limit V limit (25 m / s), the first reserved candidate box is retained as the second reserved candidate box for target A. Further, the physical projected base area S of the target is calculated systematically. phy After coordinate mapping and area calculation, S is obtained. phy=0.4 square meters. The system encapsulates the previous calculation results and extracts the physical constraint tuple of the second retained candidate box. Its internal dataset is (Attr=1,V phy =1.6,S phy =0.4).

[0059] Step S4: The system uses the vertex pixel coordinates of the second candidate bounding box of target A as the cropping boundary, generates an image tensor of dimension 3×224×224, and inputs it into a convolutional neural network. The output semantic category label of the second candidate bounding box is "personnel". The system obtains the physical limit constraints corresponding to the "personnel" label by looking up a table: reasonable region attribute set C. attr ={1}, reasonable speed range [V] min V max ] = [0, 10] m / s, reasonable base area range [S min ,S max = [0.05, 1.0] square meters. System execution logic judgment: In the extracted physical constraint tuple, the region attribute Attr=1 belongs to {1}, and the physical velocity V phy =1.6 falls within the interval [0, 10], and the physical projection base area S phy =0.4 is within the interval [0.05, 1.0]. All parameters fully satisfy all physical limit constraints. See attached... Figure 5 As shown in the figure, this diagram intuitively illustrates the spatial position verification relationship between the physical constraint tuple and the physical limit constraint condition. Figure 5 The rectangular area enclosed by the black dashed line represents the reasonable parameter boundary (area 0.05 to 1.0, velocity 0 to 10) for the semantic category label "personnel". Figure 5 In this embodiment, the scattered points marked with black "+" represent target A, whose coordinates are precisely located at (0.4, 1.6), falling entirely within the dashed box of "personnel," proving that semantic recognition and three-dimensional physical attributes are logically unified. The system confirms that target A is a real physical intrusion entity, captures on-site images, triggers an audible and visual alarm, and pushes data to the platform, completing the rapid identification of the intrusion target in the unmanned water station.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapid identification of intrusion targets in an unmanned water station, characterized in that, Includes the following steps: During the initialization phase, a static region mask matrix is ​​established to divide the water surface region and the land region. The homography matrix corresponding to the water surface region and the land region is obtained respectively, resulting in the water surface homography matrix and the land homography matrix. Based on the static region mask matrix, the water surface homography matrix and the land homography matrix are called to calculate the perspective depth and stitch them together to generate a physical depth map. Background modeling is performed on the video stream to extract initial motion candidate boxes; The initial motion candidate box is assigned a frame interval number and ratio threshold based on the physical depth map. A time window is determined based on the frame interval number. The corresponding consecutive time frames within the time window are extracted. The average motion vector between each pair of adjacent frames within the time window is calculated. Based on the horizontal and vertical components of all the average motion vectors within the time window, vector displacement synthesis and scalar path accumulation are calculated to obtain the vector time integral magnitude and absolute value integral magnitude of the average motion vector. The directional consistency ratio is obtained by dividing the vector time integral magnitude by the sum of the absolute value integral magnitude and the positive constant; the initial motion candidate boxes with a directional consistency ratio less than the ratio threshold are discarded to obtain the first retained candidate boxes; Extract the perspective reference point of the first retained candidate box; determine the region attribute to which the perspective reference point belongs, and match the corresponding homography matrix according to the region attribute; combine the average motion vector to perform inverse displacement cumulative calculation on the perspective reference point pixel coordinates under the current time frame index to obtain the perspective reference point pixel coordinates under the historical reference time frame index; use the matched homography matrix to map the perspective reference point pixel coordinates corresponding to the current time frame and the historical reference time frame to the physical surface for normalization processing to obtain the corresponding absolute horizontal physical coordinates, and then calculate the physical velocity; The physical speed is compared with the speed limit, and the first retained candidate box with the physical speed greater than the speed limit is discarded to obtain the second retained candidate box; Extract the pixel width and pixel height of the second retained candidate box and the endpoint pixel coordinates of the bottom edge; use the matched homography matrix to map the endpoint pixel coordinates to the physical surface to calculate the physical width; estimate the physical depth height based on the physical width, pixel height and pixel width, and multiply to obtain the physical projection bottom area; encapsulate the region attributes, physical velocity and physical projection bottom area to form a physical constraint tuple; The semantic category label of the second retained candidate box is output using a convolutional neural network. The physical limit constraint conditions corresponding to the semantic category label are obtained by looking up a table in a pre-configured semantic physical prior database. The physical limit constraint conditions include a reasonable set of regional attributes, a reasonable speed range, and a reasonable bottom area range. The semantic category label is verified with the physical constraint tuple and an alarm is triggered.

2. The method for rapid identification of intrusion targets in unmanned water stations according to claim 1, characterized in that, The steps for background modeling and initial motion candidate boxes in the video stream include: The decoded video stream frame sequence is obtained and grayscale processing is performed to obtain grayscale image frames; A Gaussian mixture model is used to model the background of the grayscale image frame, separate the foreground pixels, and output a binarized foreground mask image. Perform morphological opening and morphological closing operations on the binarized foreground mask image; The minimum bounding rectangle of independent connected components is calculated using a connected component analysis algorithm. The minimum bounding rectangle with a pixel area smaller than a preset lower limit is then removed to obtain the initial motion candidate box.

3. The method for rapid identification of intrusion targets in unmanned water stations according to claim 1, characterized in that, The step of allocating frame interval quantity and ratio threshold to the initial motion candidate box based on the physical depth map includes: Extract the center point coordinates of the initial motion candidate box, and obtain the depth value in the physical depth map based on the center point coordinates; Obtain the pre-defined effective depth determination range, frame interval number extreme range, and ratio threshold extreme range; The number of frame intervals corresponding to the depth value within the extreme range of the number of frame intervals is calculated using a truncated linear mapping and a direct proportional interpolation formula, and the ratio threshold corresponding to the depth value within the extreme range of the ratio threshold is calculated using an inverse proportional interpolation formula.

4. The method for rapid identification of intrusion targets in unmanned water stations according to claim 1, characterized in that, The step of using a convolutional neural network to output the semantic category label of the second retained candidate box includes: The color pixel cropping region is obtained by using the vertex pixel coordinates of the second candidate box as the cropping boundary to extract the original color image frame; The color pixel cropping region is scaled to a fixed spatial resolution using a bilinear interpolation algorithm and then numerically normalized to generate an image tensor. The image tensor is input into the convolutional neural network for forward propagation calculation to obtain a probability distribution vector. The business category corresponding to the component with the largest value in the probability distribution vector is extracted as the semantic category label.

5. The method for rapid identification of intrusion targets in an unmanned water station according to claim 1, characterized in that, The step of verifying the semantic category label with the physical constraint tuple and triggering an alarm includes: Determine whether the region attribute belongs to the reasonable region attribute set, verify whether the physical speed falls within the reasonable speed range, and determine whether the physical projection bottom area is within the reasonable bottom area range. If all the determination conditions are met, then the target corresponding to the second reserved candidate box is confirmed to be a real physical intrusion entity; The current color original image frame containing the actual physical intrusion entity is captured as a scene capture image. The semantic category label and the physical speed are extracted, and the scene capture image is pushed to the remote management platform and triggers the sound and light alarm device to issue a warning.

Images