Urban rail transit video lightweight management and linkage disposal method

Abstract

This invention provides a lightweight governance and coordinated response method for urban rail transit video, comprising: collecting urban rail transit video streams and corresponding business data; establishing a unified spatiotemporal benchmark and data mapping through preprocessing; extracting static background images through background modeling; decoupling the video in real time into static background and dynamic foreground; and constructing a ring-shaped buffer at the edge; extracting structured semantic features from the dynamic foreground to form a lightweight video stream for transmission and storage; integrating business data and feature analysis; triggering a first-level wake-up command when the warning rules are met; extracting the original foreground pixel stream T seconds before the warning from the buffer; generating key event frames using adaptive ROI encoding; matching and recording coordinated response rules based on the warning results; and finally saving the lightweight video stream, key event frames, warning results, and response records.

Description

Technical Field

[0001] This invention relates to a rail transit information management technology, and in particular to a method for lightweight management and coordinated handling of urban rail transit video. Background Technology

[0002] Rail transit video has become a crucial data foundation for station management, train operation organization, passenger flow analysis, and emergency response. Most existing urban rail transit video systems employ a model of continuous data acquisition by front-end cameras, aggregation and storage on a central platform, and manual or semi-automatic playback and analysis. The following problems exist in this model: (1) Due to the large number of cameras, high video resolution, and long running time, video data will occupy a lot of bandwidth and storage resources in the transmission and storage process, which can easily cause problems such as high load on the central platform, high network pressure, and low efficiency in retrieving historical videos; (2) In some application scenarios, although traditional video compression methods can reduce storage volume, there are cases where compression effects are obtained by reducing frames, extracting frames, or sacrificing video continuity, which can easily affect the integrity, time continuity, and subsequent event analysis effects of video data, making it difficult to meet the requirements of video evidence collection, event review, and intelligent analysis in rail transit scenarios; (3) Existing systems are usually based on collection, storage, and playback, lacking a lightweight governance mechanism for rail transit scenarios, and cannot efficiently generate key event frames, summary information, and early warning results on the basis of fully saving lightweight video data, and it is also difficult to effectively integrate with train operation status, passenger flow status, equipment status, and existing early warning rules. Therefore, there is an urgent need for a lightweight governance and coordinated response method for urban rail transit videos that can achieve lightweight transmission and full lightweight storage of videos while maintaining the video frame rate and temporal continuity, based on existing cameras and existing video surveillance systems, and further complete video analysis, key event frame and summary generation, early warning and coordinated response. Summary of the Invention

[0003] The purpose of this invention is to provide a lightweight management and coordinated handling method for urban rail transit video surveillance, comprising:

[0004] Step S100: Collect raw video streams based on existing cameras in urban rail transit and access the business data corresponding to the raw video streams;

[0005] Step S200: Preprocess the raw video stream and business data to establish a unified time reference, spatial identifier and data mapping relationship;

[0006] Step S300: Perform background modeling on the original video stream and extract the static reference background image corresponding to the spatial identifier;

[0007] Step S400: The preprocessed original video stream is decoupled into a static background region and a dynamic foreground region in real time; a circular buffer is constructed on the edge side to continuously buffer the original foreground pixel stream of the dynamic foreground region;

[0008] Step S500: Extract structured semantic features from the dynamic foreground region, and transmit and store the generated structured semantic feature vector as a lightweight video stream in real time.

[0009] Step S600: Real-time analysis of business data, and fusion analysis of business data with structured semantic feature vectors. When the analysis result meets the preset early warning rules, a first-level lightweight wake-up command is triggered and an early warning result is generated.

[0010] Step S700: In response to the first-level lightweight wake-up command, extract the original foreground pixel stream from the circular buffer from T seconds before the warning time to the current time, and enter the adaptive ROI encoding mode to generate key event frames.

[0011] Step S800: Match the corresponding linkage response rules according to the early warning result, generate linkage response instructions, and record the response record;

[0012] Step S900: Save the lightweight video stream, key event frames, early warning results, and handling records.

[0013] Furthermore, in step S200, a standardized video stream with a unified time reference and spatial identifier will be mapped to standardized business data. This specifically includes:

[0014] Calculate the mapping confidence between standardized video streams and standardized business data based on temporal overlap and spatial correlation.

[0015] When the mapping confidence level reaches a preset threshold, a unique association index number is assigned to the video data frames and business data records within the corresponding time period, thereby establishing a dynamic logical mapping relationship between video data and business data; whereby...

[0016] The process of calculating the mapping confidence score includes:

[0017] Step S231: Extract the timestamp t of the i-th frame of the normalized video stream. vi And the timestamp t of the j-th record in the standardized business data. sj Calculate the absolute value of the time difference between the two, Δt = |t vi -t sj |;

[0018] Step S232, calculate the time overlap coefficient M t ,

[0019] M t=e -λΔt ,

[0020] Where λ is the decay rate constant;

[0021] Step S233: Extract the unified spatial vector label L corresponding to the video stream. vi Business space identifier L corresponding to business data sj ; will L vi and L sj Each feature vector is decomposed into an M-dimensional spatial feature vector L=[l1,l2,l3,...,l] with hierarchical inclusion relationships. M ], where l1 to l M This includes information on route signs, station signs, area signs, and specific location signs;

[0022] Step S234, compare L level by level vi With L sj The element values ​​of the corresponding dimension in the vector determine the highest level k that is continuously and completely consistent starting from the first level. The overlap rate coefficient M is then calculated using the level overlap function. s ',

[0023] Ms'=k / M;

[0024] Step S235: For cases where the levels do not completely overlap, call the preset rail transit spatial topology map and query L. vi With L sj In the topological graph, the shortest logical path length D and the corrected spatial correlation coefficient M are used. s ,

[0025] M s =M s '·e -βD ,

[0026] Where β is the spatial decay constant;

[0027] Step S236: Calculate the mapping confidence R between the i-th video frame and the j-th service data using the weighted summation formula. ij ,

[0028] R ij =ω t ·M t +ω s ·M s ,

[0029] Where, ω t and ω s These are the time dimension weights and the space dimension weights, respectively, and ω t +ω s =1;

[0030] Step S237, calculate the mapping confidence level R. ij Compare with the preset confidence threshold γ; if R ij If ≥γ, then the i-th video frame and the j-th business data are determined to have a logical relationship. A unique association index number is assigned to the two, and the business data is encapsulated as a tag into the metadata of the corresponding video frame to complete the establishment of the data mapping relationship.

[0031] Furthermore, the specific process of step S300 includes:

[0032] Step S310: Obtain a standardized video stream with a unified spatial identifier;

[0033] Step S320: Pixel-wise sampling is performed on N consecutive frames of images in the standardized video stream using a Gaussian mixture model; by comparing the pixel values ​​of the current frame with historical distribution features, moving objects and instantaneous noise in the original video stream are identified and removed, and steady-state feature values ​​at each pixel position are extracted; the steady-state feature values ​​are spatially mapped to initially construct an original background model bound to spatial identifiers.

[0034] Step S330: Perform spatial filtering on the original background model to fill in the hole areas caused by target occlusion, and perform brightness normalization to eliminate the influence of illumination fluctuations; render a complete static reference background image without moving targets from the optimized background model; associate and encapsulate the static reference background image with the corresponding spatial identifier.

[0035] Step S340: Index and store the static reference background image and its corresponding spatial identifier in real time to build a global static reference background library.

[0036] Furthermore, the specific process of step S400 includes:

[0037] Step S410: Select the current video frame F to be processed from the standardized video stream. curr Compared with static reference background image B base Perform pixel-by-pixel alignment and background subtraction to calculate the pixel grayscale difference matrix ΔP between the two.

[0038] ΔP=|F curr -B base |;

[0039] Step S420, using dynamic threshold T diff The pixel grayscale difference matrix ΔP is binarized to identify significant change regions where the difference exceeds a threshold.

[0040] Step S430: Through connected component analysis and morphological opening and closing operations, isolated noise pixels are eliminated and holes inside the target are filled to generate a binary foreground mask image M representing the currently moving target. mask ;

[0041] Step S440, using the foreground mask image M mask For the index, the current video frame F to be processed curr Perform pixel-level segmentation, defining the pixel values ​​of the mask-covered area as the dynamic foreground region and the mask-uncovered area as the static background region;

[0042] In step S450, the pixel stream of the segmented dynamic foreground region is input into the annular buffer on the edge side in real time; at the same time, the static background region is zeroed out to stop the transmission of its physical pixel data.

[0043] Furthermore, in step S500, a YOLOv10 neural network is used to extract structured semantic features from the dynamic foreground region. The specific process includes:

[0044] Step S510: Real-time acquisition of the output dynamic foreground region pixel stream; calling the YOLOv10 neural network to scan the dynamic foreground region, identifying the dynamic target categories contained therein, and extracting the bounding box coordinate information of each dynamic target;

[0045] Step S520: For each located dynamic target, a deep feature extraction network is used to perform fine-grained semantic analysis to extract multi-dimensional structured features of each dynamic target. The multi-dimensional structured features include visual features, behavioral features, and attribute features. The visual features include the target's shape, color, and texture. The behavioral features include motion trajectory, velocity, and displacement vector. The attribute features include target type label and confidence score.

[0046] Step S530: Map the extracted multidimensional structured features, unified time reference and spatial identifier into a fixed-length structured semantic feature vector;

[0047] In step S540, the generated structured semantic feature vector is defined as a lightweight video stream. The edge ends stop transmitting the original pixel video data and only upload the lightweight video stream to the server in real time.

[0048] Furthermore, in step S520, a YOLOv10 neural network is used to obtain visual features and attribute features, and a multi-target tracking algorithm is introduced after the YOLOv10 neural network to obtain behavioral features; wherein

[0049] The calculation process for confidence scores in attribute features includes:

[0050] Step S521: The prediction head of the YOLOv10 detection network calculates the original probability distribution value P of the presence of a dynamic target within the target cell in the dynamic foreground region. obj ;

[0051] Step S522: Given that the target exists, calculate the probability value P(class) of the target belonging to a specific rail transit target category c using the Softmax function of the network output layer. c |obj);

[0052] Step S523: Calculate the final confidence score S using the multiplicative fusion formula.

[0053] S=P obj ×P(class c |obj),

[0054] Step S524: Compare the confidence score S with the business security threshold γ1, and retain the dynamic target features where S≥γ1;

[0055] The process of acquiring behavioral characteristics includes:

[0056] In step S525, the system assigns a globally unique tracking identifier to each dynamic target entering the monitoring field of view, and acquires in real time the bounding box coordinates, target type labels, and confidence scores of each dynamic target output by the YOLOv10 prediction head in the current frame, and defines them as the observation set Z at the current moment. t Simultaneously retrieve the tracking trajectory set T from the previous moment. t-1 ;

[0057] Step S526: Based on the target's motion velocity and direction in historical frames, predict its theoretical spatial position at the current moment, and generate the corresponding predicted bounding box set B. t ;

[0058] Step S527, calculate the observation set Z t The bounding boxes and the predicted bounding box set B in the data t The spatial overlap between the targets is calculated; the cosine similarity between the current observed target and the historical trajectory target on the visual feature vector is calculated using the lightweight embedding network OSNet; the spatial overlap and visual feature similarity are combined to construct the association cost matrix at the current moment.

[0059] Step S528: Solve the correlation cost matrix using the Hungarian algorithm to achieve the global optimal match between the observed target and the historical tracking trajectory at the current moment; continue the original unique tracking identifier for the successfully matched target, and assign a new unique tracking identifier to the newly appearing unmatched target;

[0060] Step S529: Extract the center point coordinate sequence of dynamic targets with the same unique tracking identifier within N consecutive frames to generate motion trajectories; combine with a unified time reference to calculate the coordinate offset of the target between adjacent frames to generate displacement vectors; calculate the ratio of the displacement vector to the corresponding time difference to obtain the real-time motion velocity of the target; standardize and encapsulate the motion trajectory, displacement vector, and real-time motion velocity to generate the behavioral features of the dynamic target.

[0061] Furthermore, step S600 specifically includes:

[0062] Step S610: Real-time acquisition of structured semantic feature vectors and standardized business data with established mapping relationships; parsing and stripping the standardized business data to separate preset early warning rule data and business status data.

[0063] Step S620: Based on a unified time base and a unified spatial identifier, extract multi-dimensional structured features from the structured semantic feature vector, and concatenate the multi-dimensional structured features with the corresponding business status data to generate fused feature data.

[0064] Step S630: Input the fused feature data into the separated preset early warning rule data for logical judgment and generate analysis results;

[0065] Step S640: When the generated analysis result meets the warning conditions, extract the globally unique tracking identifier and unified time base bound in the structured semantic feature vector that triggered the analysis result as the warning timestamp and generate the warning result; at the same time, trigger the first-level lightweight wake-up instruction and encapsulate and carry the globally unique tracking identifier and the warning timestamp into the first-level lightweight wake-up instruction.

[0066] Furthermore, step S630 specifically includes:

[0067] Step S631: Extract structured semantic feature vectors containing target object category, motion trajectory and speed from the fused feature data, as well as business data containing train operation status, passenger flow status and equipment status; call the safety threshold matrix in the early warning rule data carried in the fused feature data. The safety threshold matrix is ​​a logical judgment set composed of speed limit, prohibited entry area identifier and density threshold corresponding to different dynamic target categories.

[0068] Step S632: Compare the speed values ​​in the fused feature data with the speed upper limit values ​​in the safety threshold matrix, and calculate the speed difference between the two; if the speed difference is greater than zero, generate a first warning marker.

[0069] Step S633: Perform a point-by-point intersection operation between the sequence of pixel coordinates corresponding to the motion trajectory in the fused feature data and the set of binary pixel coordinates corresponding to the prohibited entry area marker in the safety threshold matrix; if the intersection is not empty, it is determined that the motion trajectory and the prohibited entry area have spatial overlap, and a second warning marker is generated.

[0070] Step S634: Obtain the train operation status from the business data and convert it into a status feature code; perform bit-by-bit matching of the status feature code with the preset linkage status logic rules in the safety threshold matrix;

[0071] Step S635: The matching results of the first warning marker, the second warning marker, and the status feature code are weighted and summed to obtain the rule matching score; if the rule matching score reaches the preset logical threshold, the current scene is determined to meet the warning conditions, and the analysis result is generated.

[0072] Furthermore, the specific process of step S700 includes:

[0073] In step S710, the system retrieves the corresponding original image sequence from the circular buffer on the edge side based on the globally unique tracking identifier and warning timestamp carried in the first-level lightweight wake-up instruction; and locates and extracts the original foreground pixel stream from T seconds before the warning time to the current time.

[0074] Step S720: Enter the adaptive ROI encoding mode, define the specific pixel area containing the warning target and its motion trajectory as the high-quality ROI area, and define the rest of the image except for the high-quality ROI area as the low-quality non-ROI area.

[0075] Step S730: Perform non-uniform quantization encoding processing on the extracted pixel stream and assign quantization parameters QP to the high-quality ROI regions. roi To preserve the visual detail features of the target; assign quantization parameters QP to low-optimal non-ROI regions. non-roi QP reduces bitrate by drastically discarding redundant background information. non-roi >QP roi ;

[0076] In step S740, the encoded lightweight foreground video stream is transmitted back to the server via the network. The server uses the spatial identifier and temporal reference in the transmitted stream to retrieve the corresponding static reference background image, executes pixel-level background restoration and fusion algorithms, and finally generates key event frames with complete environmental semantics.

[0077] Furthermore, the specific process of step S800 includes:

[0078] Step S810: Obtain the early warning results in real time and parse out the corresponding target type label, early warning time, and unified spatial identifier;

[0079] Step S820: Based on the parsed target type label and unified spatial identifier, retrieve the preset action instruction mapping table;

[0080] Step S830: Extract the handling action that matches the warning result from the action instruction mapping table, and combine it with the unified spatial identifier to locate the corresponding execution terminal, and generate a linkage handling instruction that includes the execution target, execution action and execution sequence;

[0081] In step S840, the generated linkage handling command is output to the corresponding execution terminal, and the execution status is recorded synchronously.

[0082] Compared with the prior art, the present invention has the following advantages: (1) The present invention splits the video into static background and dynamic foreground in real time. Only a very small volume of structured semantic feature vector is uploaded at the edge as a lightweight stream, realizing lightweight storage of the full data and greatly reducing bandwidth occupation; (2) The present invention has a ring buffer on the edge side, which continuously caches the original foreground pixel stream for T seconds. When an early warning is triggered, the system only encodes the high-quality ROI area (early warning target) with high quality and combines it with the stored background image to synthesize key event frames. While maintaining high frame rate and continuity, it ensures the clarity of key evidence; (3) The present invention establishes a unified time and space benchmark and logically maps the video stream with business data such as train operation, passenger flow, and equipment status. Attached Figure Description

[0083] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0084] Figure 2 This is a schematic diagram of a method flow according to an embodiment of the present invention. Detailed Implementation

[0085] A lightweight management and coordinated response method for urban rail transit video surveillance includes the following steps:

[0086] Step S100: Collect raw video streams based on existing cameras in urban rail transit and access business data corresponding to the raw video streams. The business data includes at least one or more of the following: train operation status data, passenger flow status data, equipment status data, and preset early warning rule data.

[0087] Step S200: Preprocess the raw video stream and business data to establish a unified time reference, spatial identifier and data mapping relationship;

[0088] Step S300: Perform background modeling on the original video stream and extract the static reference background image corresponding to the spatial identifier;

[0089] Step S400: Decouple the preprocessed original video stream into a static background region and a dynamic foreground region in real time; construct a ring buffer on the edge side to continuously buffer the original foreground pixel stream of the dynamic foreground region for T seconds;

[0090] Step S500: Extract structured semantic features from the dynamic foreground region, and transmit and store the generated structured semantic feature vector as a lightweight video stream in real time.

[0091] Step S600: Real-time analysis of business data, and fusion analysis of business data with structured semantic feature vectors. When the analysis result meets the preset early warning rules, a first-level lightweight wake-up command is triggered and an early warning result is generated.

[0092] Step S700: In response to the first-level lightweight wake-up instruction, extract the original foreground pixel stream from the circular buffer from T seconds before the warning time to the current time, and enter the adaptive ROI encoding mode: (1) Dynamically delineate the high-optimal ROI region and low-optimal non-ROI region in the original foreground pixel stream based on the business data; (2) Assign quantization parameter QP to the high-optimal ROI region. roi Quantization parameter QP is assigned to low-optimal non-ROI regions. non-roi Differential coding is performed, where QP roi <QP non-roi (3) The encoded foreground video stream is transmitted back and synthesized with the static reference background image stored on the server to generate key event frames;

[0093] Step S800: Match the corresponding linkage response rules according to the early warning result, generate linkage response instructions, and record the response record;

[0094] Step S900: Save the lightweight video stream, key event frames, early warning results, and handling records.

[0095] The process of establishing a unified time base in step S200 includes:

[0096] Obtain a pre-set standard time source within the urban rail transit system as a unified time reference; obtain the original video timestamps of the original video stream and the original business timestamps of each dimension of business data; calculate the time offset of the original video timestamps and each original business timestamp relative to the unified time reference; use the time offsets to perform time axis translation correction on the original video stream and business data respectively, and generate standardized video streams and standardized business data with a unified time reference.

[0097] The process of establishing a unified spatial identifier in step S200 includes: obtaining the physical installation location information of the camera to which the original video stream belongs, and retrieving the rail transit spatial topology model associated with its geographical location; based on a preset spatial mapping function, mapping the physical installation location information into a unified spatial vector label containing station identifiers, area identifiers, and specific location identifiers; at the same time, formatting and cleaning the location fields carried in the business data, converting them into business spatial identifiers that are logically consistent with the unified spatial vector label, thereby achieving semantic unity between video and business data in the spatial dimension.

[0098] The process of establishing a data mapping relationship in step S200 includes: using a unified time base and a unified spatial identifier as retrieval dimensions, placing standardized video streams and standardized business data within a preset time sliding window for spatiotemporal matching; calculating the mapping confidence between the two based on the time overlap and spatial correlation; and when the mapping confidence reaches a preset threshold, assigning a unique association index number to the video data frames and business data records within the corresponding time period, thereby establishing a dynamic logical mapping relationship between video data and business data.

[0099] In step S200, establishing the data mapping relationship involves calculating the mapping confidence between time overlap and spatial correlation. The specific process includes:

[0100] Step S231: Extract the timestamp t of the i-th frame of the normalized video stream. vi And the timestamp t of the j-th record in the standardized business data. sj Calculate the absolute value of the time difference between the two, Δt = |t vi -t sj |;

[0101] Step S232, calculate the time overlap coefficient M t ,

[0102] M t =e -λΔt ,

[0103] Where λ is the decay rate constant;

[0104] Step S233: Extract the unified spatial vector label L corresponding to the video stream. vi Business space identifier L corresponding to business data sj ; will L vi and L sj Each feature vector is decomposed into an M-dimensional spatial feature vector L=[l1,l2,l3,...,l] with hierarchical inclusion relationships. M ], where l1 to l M These represent route signs, station signs, area signs (such as platform / concourse / passage), and specific location signs;

[0105] Step S234, compare L level by level vi With L sj The element values ​​of the corresponding dimension in the vector determine the highest level k that is continuously and completely consistent starting from the first level (line). The overlap rate coefficient M is calculated using the level overlap function. s ',

[0106] Ms'=k / M,

[0107] For example, if L vi For "Metro Line 1 - a certain station - platform - camera number 01", L sj If the line is "Metro Line 1 - a certain station - turnstile", then it overlaps at both the line and station levels. If the total number of levels M=4, then the level overlap rate M is... s =0.5;

[0108] Step S235: For cases where the levels do not completely overlap, call the preset rail transit spatial topology map and query L. vi With L sj The shortest logical path length D (i.e., the minimum number of hops across functional areas) in the topology diagram is used to adjust the spatial correlation coefficient M. s ,

[0109] M s =M s '·e -βD ,

[0110] Where β is the spatial decay constant;

[0111] Step S236: Calculate the mapping confidence R between the i-th video frame and the j-th service data using the weighted summation formula. ij ,

[0112] R ij =ω t ·M t +ω s ·M s ,

[0113] Where, ω t and ω s These are the time dimension weights and the space dimension weights, respectively, and ω t +ω s =1;

[0114] Step S237, calculate the mapping confidence level R. ij Compare with the preset confidence threshold γ; if R ijIf ≥γ, then the i-th video frame and the j-th business data are determined to have a logical relationship. A unique association index number is assigned to the two, and the business data is encapsulated as a tag into the metadata of the corresponding video frame to complete the establishment of the data mapping relationship.

[0115] The specific process of step S300 includes:

[0116] Step S310: Obtain a standardized video stream with a unified spatial identifier; retrieve the scene attributes (such as platform, tunnel or station hall) of the video acquisition point based on the spatial identifier; initialize background modeling parameters based on the scene attributes, including the learning rate α and the background model update threshold T, to establish an initial modeling environment for a specific spatial point.

[0117] Step S320: Pixel-by-pixel sampling is performed on N consecutive frames of images in the standardized video stream using a Gaussian mixture model (GMM); by comparing the pixel values ​​of the current frame with historical distribution features, moving objects and instantaneous noise in the original video stream are identified and removed, and steady-state feature values ​​at each pixel location are extracted; the steady-state feature values ​​are spatially mapped to initially construct an original background model bound to spatial identifiers; wherein the steady-state feature values ​​refer to the probability model composed of K Gaussian distributions established for each pixel.

[0118] Step S330: Perform spatial filtering on the original background model to fill in the hole areas caused by target occlusion, and perform brightness normalization to eliminate the influence of illumination fluctuations; render a complete static reference background image without moving targets from the optimized background model; associate and encapsulate the static reference background image with the corresponding spatial identifier as the core reference frame for subsequent video lightweighting processing.

[0119] In step S340, the static reference background image and its corresponding spatial identifier are indexed and stored in real time to build a global static reference background library. At the same time, the edge end repeats steps S310 to S330 according to the update cycle to realize dynamic calibration and version iteration of the static reference background image.

[0120] The specific process of step S400 includes:

[0121] Step S410: Obtain the standardized video stream and the corresponding spatial identifier static reference background image; transfer the current video frame F to be processed in the standardized video stream. curr Compared with static reference background image B base Perform pixel-by-pixel alignment and background subtraction to calculate the pixel grayscale difference matrix ΔP between the two.

[0122] ΔP=|F curr -B base |;

[0123] Step S420, using dynamic threshold T diff The pixel grayscale difference matrix ΔP is binarized to identify significant change regions where the difference exceeds a threshold.

[0124] Step S430: Through connected component analysis and morphological opening and closing operations, isolated noise pixels are eliminated and holes inside the target are filled to generate a binary foreground mask image M representing the currently moving target. mask ;

[0125] Step S440, using the foreground mask image M mask For the index, the current video frame F to be processed curr Perform pixel-level segmentation; define the pixel values ​​of the mask-covered area as the dynamic foreground region, and retain its original image features for feature extraction; define the mask-uncovered area as the static background region, and determine that the pixels in this region have logical consistency with the static reference background image;

[0126] In step S450, the segmented dynamic foreground region pixel stream is input into the ring buffer on the edge side in real time for subsequent feature extraction and T-second raw backtracking buffering; at the same time, the static background region is zeroed out to stop the transmission of its physical pixel data, thereby realizing real-time decoupling and data compression of the video stream at the semantic level.

[0127] In step S500, a YOLOv10 neural network is used to extract structured semantic features from the dynamic foreground region. The specific process includes:

[0128] Step S510: Real-time acquisition of the output dynamic foreground region pixel stream; calling the YOLOv10 neural network to scan the dynamic foreground region, identifying the dynamic target categories contained therein, extracting the bounding box coordinate information of each dynamic target, and realizing accurate spatial positioning of the dynamic targets; the dynamic target categories include at least pedestrians, luggage, abnormally moving objects, and track obstacles;

[0129] Step S520: For each located dynamic target, fine-grained semantic analysis is performed using a deep feature extraction network to extract multi-dimensional structured features of each dynamic target. The multi-dimensional structured features include visual features, behavioral features, and attribute features. Visual features include target shape, color, and texture. Behavioral features include motion trajectory, speed, and displacement vector. Attribute features include target type label and confidence score.

[0130] Step S530: The extracted multidimensional structured features are aggregated according to a preset metadata protocol and combined with the generated unified time reference and spatial identifier to map into a fixed-length structured semantic feature vector.

[0131] In step S540, the generated structured semantic feature vector is defined as a lightweight video stream. The edge device stops transmitting the original pixel video data and only uploads the lightweight video stream to the server in real time. After receiving the lightweight video stream, the server writes it into the structured database for full storage to support subsequent abnormal event retrieval and business linkage analysis.

[0132] In step S510, the pixel stream of the dynamic foreground region is input into the preset YOLOv10 target detection network. The multi-scale spatial features of the foreground pixels are extracted through the backbone network with large kernel convolution of the YOLOv10 network, and the multi-scale spatial features are fused and enhanced by the path aggregation network (PAN) to generate a fused feature map containing semantic and positional information. The fused feature map is input into the decoupled detection head of YOLOv10, and end-to-end prediction is performed using its consistent dual assignment strategy: (1) The classification branch predicts the probability distribution of each pixel region belonging to the preset dynamic target category based on the feature map to obtain the target type label of each dynamic target; (2) The regression branch calculates the offset of each dynamic target relative to the image coordinate system using the anchorless prediction mechanism to generate the bounding box coordinate information of each dynamic target. The target type label and bounding box coordinate information of each dynamic target are associated and encapsulated, and combined with a unified time base and spatial identifier, a globally unique target sequence number is assigned to each identified dynamic target to generate a structured preliminary target detection result.

[0133] In step S520, a YOLOv10 neural network is used to obtain visual and attribute features. A multi-object tracking algorithm is then introduced after the YOLOv10 neural network to obtain behavioral features. The confidence score calculation process includes:

[0134] Step S521: The prediction head of the YOLOv10 detection network performs feature activation on the target unit (pixel) in the dynamic foreground region, and calculates the original probability distribution value P of the presence of a dynamic target in the unit. obj This is used to characterize the certainty that the spatial location belongs to a foreground target rather than background noise;

[0135] Step S522: Given that the target exists, the probability value P(class) of the target belonging to a specific rail transit target category c (such as pedestrians, luggage, etc.) is calculated using the Softmax function of the network output layer. c |obj);

[0136] Step S523: Calculate the final confidence score S using the multiplicative fusion formula.

[0137] S=P obj ×P(class c|obj),

[0138] Step S524: Compare the confidence score S with the business security threshold γ1, and retain only the dynamic target features that S≥γ1, thereby eliminating false target recognition results caused by environmental interference such as lighting and shadows.

[0139] The process of obtaining behavioral features in step S520 includes:

[0140] In step S525, the system assigns a globally unique tracking identifier to each dynamic target entering the monitoring field of view, and acquires in real time the bounding box coordinates, target type labels, and confidence scores of each dynamic target output by the YOLOv10 prediction head in the current frame, and defines them as the observation set Z at the current moment. t Simultaneously retrieve the tracking trajectory set T from the previous moment. t-1 ;

[0141] Step S526: Use the Kalman filter algorithm to process the set of survival tracking trajectories T. t-1 State extrapolation is performed, and based on the target's velocity and direction in historical frames, its theoretical spatial position at the current moment is predicted, generating a corresponding set of predicted bounding boxes B. t ;

[0142] Step S527, calculate the observation set Z t The bounding boxes and the predicted bounding box set B in the data t The spatial overlap between the observed targets and historical trajectory targets is calculated using the lightweight embedded network OSNet. The cosine similarity between the observed targets and historical trajectory targets on the visual feature vectors is then used to calculate the spatial overlap and visual feature similarity to construct the association cost matrix at the current moment.

[0143] Step S528: Solve the correlation cost matrix using the Hungarian algorithm to achieve the global optimal match between the observed target and the historical tracking trajectory at the current moment; continue the original unique tracking identifier for the successfully matched target, and assign a new unique tracking identifier to the newly appearing unmatched target;

[0144] Step S529: Extract the center point coordinate sequence of dynamic targets with the same unique tracking identifier within N consecutive frames to generate motion trajectories; combine with a unified time reference to calculate the coordinate offset of the target between adjacent frames to generate displacement vectors; further calculate the ratio of the displacement vector to the corresponding time difference to obtain the real-time motion velocity of the target; standardize and encapsulate the motion trajectory, displacement vector, and real-time motion velocity to generate the behavioral features of the dynamic target.

[0145] The spatial overlap in step S527 is obtained through the intersection-over-union (IoU) ratio. In step S527, cosine similarity is calculated to obtain the visual feature vector composed of shape, color, and texture. The method for constructing the association cost matrix in step S527 includes mapping the calculated spatial overlap to a spatial cost factor IoU, and mapping the visual feature cosine similarity to a visual cost factor S. cos Ensure that all indicators are on the same order of magnitude; according to the formula C=λ·(1-IoU)+(1-λ)·(1-S cos Calculate the comprehensive association cost C for each pair of observed targets and historical trajectories, where λ is the spatial weight coefficient; create an empty matrix of dimension M×N, where M is the number of historical trajectories and N is the number of current detections, and fill the calculated comprehensive association cost into the corresponding matrix cells in sequence.

[0146] The specific process of step S530 includes:

[0147] Step S531: Obtain the multidimensional structured features of each dynamic target, normalize the multidimensional structured feature data of different dimensions and scales according to the preset metadata protocol, and reconstruct the fields according to the fixed bit width.

[0148] Step S532: Extract the unified time reference and spatial identifier bound to the unique tracking identifier of the current dynamic target; use a mapping algorithm to transform the above spatiotemporal attributes into the spatiotemporal semantic components of the dynamic target, and inject them as key context information into the aforementioned converged feature set.

[0149] Step S533: Concatenate the standardized attribute components, visual components, behavioral components, and auxiliary spatiotemporal semantic components; convert the variable-length feature description into a globally unified fixed-length structured semantic feature vector using a length padding algorithm.

[0150] Step S534: Generate a fixed-length summary information from the generated fixed-length structured semantic feature vector using an encrypted hash algorithm to ensure the integrity of the data during transmission between the edge and the cloud; output the encapsulated vector to the subsequent lightweight transmission process to achieve the final conversion of the target from image pixels to structured semantics.

[0151] Step S600 specifically includes:

[0152] Step S610: Real-time acquisition of structured semantic feature vectors and standardized business data with established mapping relationships; parsing and stripping the standardized business data to separate preset early warning rule data and business status data. Business status data includes train operation status data, passenger flow status data and equipment status data, etc.

[0153] Step S620: Based on a unified time base and a unified spatial identifier, extract multi-dimensional structured features from the structured semantic feature vector, and concatenate the multi-dimensional structured features with the corresponding business status data to generate fused feature data.

[0154] Step S630: The fused feature data is used as the input for judgment and input into the separated preset early warning rule data for logical judgment to generate analysis results;

[0155] Step S640: When the generated analysis result meets the warning conditions, extract the globally unique tracking identifier and unified time base bound in the structured semantic feature vector that triggered the analysis result as the warning timestamp and generate the warning result; at the same time, trigger the first-level lightweight wake-up instruction and encapsulate and carry the globally unique tracking identifier and the warning timestamp into the first-level lightweight wake-up instruction.

[0156] Step S630 specifically includes:

[0157] Step S631: Extract structured semantic feature vectors containing target object category, motion trajectory and speed, as well as business data containing train operation status, passenger flow status and equipment status from the fused feature data;

[0158] Step S632: Call the preset early warning rule data carried in the fused feature data; the preset early warning rule data contains a safety threshold matrix for each dynamic target category. The safety threshold matrix refers to a logical judgment set composed of the speed limit, prohibited area sign and density threshold corresponding to different dynamic target categories. It is obtained by standardizing and extracting the critical parameters in historical accident cases.

[0159] Step S633: Compare the real-time speed and motion trajectory in the structured semantic feature vector with the speed limit and prohibited entry area of ​​the corresponding dynamic target category in the safety threshold matrix and perform spatial intersection calculation; map and match the train operation status in the business data with the linkage status logic in the safety threshold matrix.

[0160] Step S634: Summarize the comparison results of each dimension in step S633, and calculate the rule matching score. The rule matching score refers to the quantitative score of the deviation between real-time feature data and preset rule data. It is obtained by normalizing the comparison deviation values ​​of each item through a weighted summation formula. If the rule matching score reaches the preset logical threshold, it is determined that the current scene meets the warning conditions and the analysis result is generated.

[0161] The safety threshold matrix in step S632 is a two-dimensional logical array based on dynamic target categories and spatial point identifiers, which can be represented as Q=[T i ,L j ]→{Vmax A mask ,ρ},T i L represents the i-th type of dynamic target. j Representing the j-th spatial point identifier, the mapping cell determined by the intersection of the two stores a set of safety judgment indicators under that spatiotemporal context. Each matrix cell encapsulates at least quantitative judgment parameters such as the speed limit, the prohibited area identifier, and the density threshold. Speed ​​limit V max : The maximum permissible operating speed for a specific target within a specific area; Prohibited area marker A mask : The set of binary coordinates corresponding to a specific target that is prohibited from appearing in the video footage at this location (e.g., pedestrians are prohibited from entering the center area of ​​the track); Density threshold ρ: The maximum number of similar dynamic targets allowed to appear within a specific area. The safety threshold matrix is ​​obtained through historical early warning data of urban rail transit. The historical early warning data includes the types of risk events that have occurred in the past, the motion parameters of the dynamic targets at the time of the event, and the corresponding business status parameters; cluster analysis is performed on the historical early warning data to extract the critical feature values ​​of the judgment parameters for each category of dynamic targets when the risk is triggered.

[0162] Step S633 specifically includes the following process:

[0163] Step S6331: Extract real-time motion speed, motion trajectory, speed limit and prohibited area markers;

[0164] Step S6332: Compare the real-time motion speed value with the speed limit value and calculate the speed difference between the two; if the speed difference is greater than zero, generate a first warning flag.

[0165] Step S6333: Perform a point-by-point intersection operation between the pixel coordinate sequence corresponding to the motion trajectory and the binary pixel coordinate set corresponding to the prohibited entry area marker; if the intersection is not empty, it is determined that the motion trajectory and the prohibited entry area have spatial overlap, and a second warning mark is generated.

[0166] Step S6334: Obtain the train operation status from the business data and convert it into a status feature code. The status feature code is a binary string generated by digitally encoding the train door opening / closing status and station entry / exit status, which is obtained by bitwise operations on the original business data. The status feature code is matched bit by bit with the preset linkage status logic rules in the safety threshold matrix.

[0167] Step S6335: Summarize the matching results of the first warning marker, the second warning marker, and the status feature code to generate a comprehensive logical judgment vector, which serves as the input data for generating the analysis results in step S634.

[0168] The specific process of step S700 includes:

[0169] In step S710, the system retrieves the corresponding original image sequence from the circular buffer on the edge side based on the globally unique tracking identifier and warning timestamp carried in the first-level lightweight wake-up instruction; locates and extracts the original foreground pixel stream from T seconds before the warning time to the current time, ensuring that the video evidence covers the complete evolution process of the risk occurrence;

[0170] Step S720: Enter the adaptive ROI encoding mode, define the specific pixel area containing the warning target and its motion trajectory as the high-quality ROI area, and define the rest of the image except for the high-quality ROI area as the low-quality non-ROI area.

[0171] Step S730 involves performing non-uniform quantization encoding processing on the extracted pixel stream, specifically including:

[0172] Assign a smaller quantization parameter QP to the high ROI region. roi To preserve the visual details of the target;

[0173] Assign a larger quantization parameter QP to the low-optimal non-ROI region. non-roi QP reduces bitrate by drastically discarding redundant background information. roi ∈[10,28], QP non-roi ∈[32,45], and QP non-roi -QP roi ≥8;

[0174] In step S740, the encoded lightweight foreground video stream is transmitted back to the server via the network. The server uses the spatial identifier and temporal reference in the transmitted stream to retrieve the corresponding static reference background image, executes pixel-level background restoration and fusion algorithms, and finally generates key event frames with complete environmental semantics.

[0175] The specific process of step S800 includes:

[0176] Step S810: Obtain the early warning results in real time and parse out the corresponding target type label, early warning time, and unified spatial identifier;

[0177] Step S820: Based on the parsed target type label and unified spatial identifier, retrieve the preset linkage response rule base; The linkage response rule base contains pre-stored action instruction mapping tables for different dynamic target categories and spatial locations. The action instruction mapping table is a logical association matrix composed of warning level, response action and execution terminal identifier, which is obtained by digitally translating the preset emergency plan.

[0178] Step S830: Extract the handling action that matches the warning result from the action instruction mapping table, and combine it with the unified spatial identifier to locate the corresponding execution terminal, and generate a linkage handling instruction that includes the execution target, execution action and execution sequence;

[0179] In step S840, the linkage handling instruction generated in step S830 is output to the corresponding execution terminal, and the execution status is recorded synchronously for unified management of the handling records in step S900.

[0180] Example 1, combined with Figure 2 This embodiment is applied to the scenario of identifying and verifying abnormal use of subway preferential cards. The specific implementation process is as follows:

[0181] Step S1, Raw Data Access: Based on the existing cameras in the urban rail transit system, raw video streams are collected from locations such as gate areas and passages, and business data corresponding to the raw video streams are accessed. This business data includes at least preferential card swipe records, station entry and exit records, train status, and preset warning rule data.

[0182] Step S2, Spatiotemporal Preprocessing and Data Mapping: The original video stream and business data are preprocessed to establish a unified time base, spatial identifier, and data mapping relationship. Specifically, the mapping confidence level is obtained by calculating the temporal overlap and spatial correlation between the standardized video stream and preferential card business data (such as card swipe records). When the mapping confidence level reaches a preset threshold, a unique association index number is assigned to the video data frames and business data records within the corresponding time period, thereby establishing a dynamic logical mapping relationship between video data and business data.

[0183] Step S3, Static reference background modeling: Perform background modeling on the original video stream, use a Gaussian mixture model to remove moving objects such as passing passengers and instantaneous noise, extract static reference background images corresponding to spatial identifiers, and construct a global static reference background library.

[0184] Step S4, real-time decoupling and caching of video stream: The preprocessed original video stream is decoupled in real time into a static background region and a dynamic foreground region (such as passing passengers); at the same time, a ring-shaped buffer is built on the edge side to continuously cache the original foreground pixel stream of the dynamic foreground region.

[0185] Step S5, Feature Extraction and Lightweight Transmission: A YOLOv10 neural network is used to extract structured semantic features from the dynamic foreground region, obtaining the target's visual, behavioral, and attribute features, which are then mapped to a fixed-length structured semantic feature vector. The generated structured semantic feature vector is transmitted and stored in real-time as a lightweight video stream, while transmission of the original pixel video data stops at the edge.

[0186] Step S6, Business Data Fusion and Early Warning Trigger: Real-time analysis of preferential card business data, and fusion analysis of business data with structured semantic feature vectors. When the analysis results (such as the consistency results of preferential card swiping records and extracted passenger behavior and attribute features) meet the preset early warning rules (determined as suspected abnormal use of preferential cards), a first-level lightweight wake-up command is triggered, and the globally unique tracking identifier and early warning timestamp are encapsulated in it to generate an early warning result.

[0187] Step S7, Adaptive Encoding and Keyframe Generation: In response to the Level 1 lightweight wake-up command, the system extracts the original foreground pixel stream from the circular buffer on the edge side, from T seconds before the warning time to the current time, and enters the adaptive ROI encoding mode. The region containing the abnormally used preferential card target and its motion trajectory is designated as a high-optimal ROI region and assigned a smaller quantization parameter QP. roi The remaining portion is designated as a low-optimal non-ROI region and assigned a larger quantization parameter QP. non-roi The encoded foreground video stream is sent back to the server and pixel-level fused with the static baseline background image to generate key event frames with complete environmental semantics.

[0188] Step S8, generating linkage handling instructions: Based on the early warning result of abnormal use of preferential cards, match the corresponding action instruction mapping table in the linkage handling rule base, generate linkage handling instructions (such as pushing alarms to the station management terminal, on-site verification, etc.), output to the corresponding execution terminal and record the handling record synchronously.

[0189] Step S9, Full Data Storage and Management: The lightweight video streams, key event frames, early warning results and handling records generated above are uniformly stored and managed throughout their lifecycle to meet the requirements of subsequent manual review and traceability.

Claims

1. A lightweight management and coordinated handling method for urban rail transit video surveillance, characterized in that, include: Step S100: Collect raw video streams based on existing cameras in urban rail transit and access the business data corresponding to the raw video streams; Step S200: Preprocess the raw video stream and business data to establish a unified time reference, spatial identifier and data mapping relationship; Step S300: Perform background modeling on the original video stream and extract the static reference background image corresponding to the spatial identifier; Step S400: Decouple the preprocessed original video stream into a static background region and a dynamic foreground region in real time; A circular buffer is constructed on the edge side to continuously buffer the original foreground pixel stream of the dynamic foreground region; Step S500: Extract structured semantic features from the dynamic foreground region, and transmit and store the generated structured semantic feature vector as a lightweight video stream in real time. Step S600: Real-time analysis of business data, and fusion analysis of business data with structured semantic feature vectors. When the analysis result meets the preset early warning rules, a first-level lightweight wake-up command is triggered and an early warning result is generated. Step S700: In response to the first-level lightweight wake-up command, extract the original foreground pixel stream from the circular buffer from T seconds before the warning time to the current time, and enter the adaptive ROI encoding mode to generate key event frames. Step S800: Match the corresponding linkage response rules according to the early warning result, generate linkage response instructions, and record the response record; Step S900: Save the lightweight video stream, key event frames, early warning results, and handling records.

2. The method according to claim 1, characterized in that, In step S200, standardized video streams with unified time references and spatial identifiers will be mapped to standardized business data. This specifically includes: Calculate the mapping confidence between standardized video streams and standardized business data based on temporal overlap and spatial correlation. When the mapping confidence level reaches a preset threshold, a unique association index number is assigned to the video data frames and business data records within the corresponding time period, thereby establishing a dynamic logical mapping relationship between video data and business data; whereby... The process of calculating the mapping confidence score includes: Step S231: Extract the timestamp t of the i-th frame of the normalized video stream. vi And the timestamp t of the j-th record in the standardized business data. sj Calculate the absolute value of the time difference between the two, Δt = |t vi -t sj |; Step S232, calculate the time overlap coefficient M t , M t =e -λΔt , Where λ is the decay rate constant; Step S233: Extract the unified spatial vector label L corresponding to the video stream. vi Business space identifier L corresponding to business data sj ; will L vi and L sj Each feature vector is decomposed into an M-dimensional spatial feature vector L=[l1,l2,l3,...,l] with hierarchical inclusion relationships. M ], where l1 to l M This includes information on route signs, station signs, area signs, and specific location signs; Step S234, compare L level by level vi With L sj The element values ​​of the corresponding dimension in the vector determine the highest level k that is continuously and completely consistent starting from the first level. The overlap rate coefficient M is then calculated using the level overlap function. s ', Ms'=k / M; Step S235: For cases where the levels do not completely overlap, call the preset rail transit spatial topology map and query L. vi With L sj In the topological graph, the shortest logical path length D and the corrected spatial correlation coefficient M are used. s , M s =M s '·e -βD , Where β is the spatial decay constant; Step S236: Calculate the mapping confidence R between the i-th video frame and the j-th service data using the weighted summation formula. ij , R ij =ω t ·M t +oh s ·M s , Where, ω t and ω s These are the time dimension weights and the space dimension weights, respectively, and ω t +ω s =1; Step S237, calculate the mapping confidence level R. ij Compare with the preset confidence threshold γ; if R ij If ≥γ, then the i-th video frame and the j-th business data are determined to have a logical relationship. A unique association index number is assigned to the two, and the business data is encapsulated as a tag into the metadata of the corresponding video frame to complete the establishment of the data mapping relationship.

3. The method according to claim 2, characterized in that, The specific process of step S300 includes: Step S310: Obtain a standardized video stream with a unified spatial identifier; Step S320: Pixel-wise sampling is performed on N consecutive frames of images in the standardized video stream using a Gaussian mixture model; by comparing the pixel values ​​of the current frame with historical distribution features, moving objects and instantaneous noise in the original video stream are identified and removed, and steady-state feature values ​​at each pixel position are extracted; the steady-state feature values ​​are spatially mapped to initially construct an original background model bound to spatial identifiers. Step S330: Perform spatial filtering on the original background model to fill in the hole areas caused by target occlusion, and perform brightness normalization to eliminate the influence of illumination fluctuations; render a complete static reference background image without moving targets from the optimized background model; associate and encapsulate the static reference background image with the corresponding spatial identifier. Step S340: Index and store the static reference background image and its corresponding spatial identifier in real time to build a global static reference background library.

4. The method according to claim 3, characterized in that, The specific process of step S400 includes: Step S410: Select the current video frame F to be processed from the standardized video stream. curr Compared with static reference background image B base Perform pixel-by-pixel alignment and background subtraction to calculate the pixel grayscale difference matrix ΔP between the two. ΔP=|F curr -B base |; Step S420, using dynamic threshold T diff The pixel grayscale difference matrix ΔP is binarized to identify significant change regions where the difference exceeds a threshold. Step S430: Through connected component analysis and morphological opening and closing operations, isolated noise pixels are eliminated and holes inside the target are filled to generate a binarized foreground mask image M representing the currently moving target. mask ; Step S440, using the foreground mask image M mask For the index, the current video frame F to be processed curr Perform pixel-level segmentation, defining the pixel values ​​of the mask-covered area as the dynamic foreground region and the mask-uncovered area as the static background region; In step S450, the pixel stream of the segmented dynamic foreground region is input into the annular buffer on the edge side in real time; at the same time, the static background region is zeroed out to stop the transmission of its physical pixel data.

5. The method according to claim 4, characterized in that, In step S500, a YOLOv10 neural network is used to extract structured semantic features from the dynamic foreground region. The specific process includes: Step S510: Real-time acquisition of the output dynamic foreground region pixel stream; calling the YOLOv10 neural network to scan the dynamic foreground region, identifying the dynamic target categories contained therein, and extracting the bounding box coordinate information of each dynamic target; Step S520: For each located dynamic target, a deep feature extraction network is used to perform fine-grained semantic analysis to extract multi-dimensional structured features of each dynamic target. The multi-dimensional structured features include visual features, behavioral features, and attribute features. The visual features include the target's shape, color, and texture. The behavioral features include motion trajectory, velocity, and displacement vector. The attribute features include target type label and confidence score. Step S530: Map the extracted multidimensional structured features, unified time reference and spatial identifier into a fixed-length structured semantic feature vector; In step S540, the generated structured semantic feature vector is defined as a lightweight video stream. The edge ends stop transmitting the original pixel video data and only upload the lightweight video stream to the server in real time.

6. The method according to claim 5, characterized in that, In step S520, a YOLOv10 neural network is used to obtain visual features and attribute features, and a multi-target tracking algorithm is introduced after the YOLOv10 neural network to obtain behavioral features. in The calculation process for confidence scores in attribute features includes: Step S521: The prediction head of the YOLOv10 detection network calculates the original probability distribution value P of the presence of a dynamic target within the target cell in the dynamic foreground region. obj ; Step S522: Given that the target exists, calculate the probability value P(class) of the target belonging to a specific rail transit target category c using the Softmax function of the network output layer. c |obj); Step S523: Calculate the final confidence score S using the multiplicative fusion formula. S=P obj ×P(class c |obj), Step S524: Compare the confidence score S with the business security threshold γ1, and retain the dynamic target features where S≥γ1; The process of acquiring behavioral characteristics includes: In step S525, the system assigns a globally unique tracking identifier to each dynamic target entering the monitoring field of view, and acquires in real time the bounding box coordinates, target type labels, and confidence scores of each dynamic target output by the YOLOv10 prediction head in the current frame, and defines them as the observation set Z at the current moment. t Simultaneously retrieve the tracking trajectory set T from the previous moment. t-1 ; Step S526: Based on the target's motion velocity and direction in historical frames, predict its theoretical spatial position at the current moment, and generate the corresponding predicted bounding box set B. t ; Step S527, calculate the observation set Z t The bounding boxes and the predicted bounding box set B in the data t The spatial overlap between the targets is calculated; the cosine similarity between the current observed target and the historical trajectory target on the visual feature vector is calculated using the lightweight embedding network OSNet; the spatial overlap and visual feature similarity are combined to construct the association cost matrix at the current moment. Step S528: Solve the correlation cost matrix using the Hungarian algorithm to achieve the global optimal match between the observed target and the historical tracking trajectory at the current moment; continue the original unique tracking identifier for the successfully matched target, and assign a new unique tracking identifier to the newly appearing unmatched target; Step S529: Extract the center point coordinate sequence of dynamic targets with the same unique tracking identifier within N consecutive frames to generate motion trajectories; combine with a unified time reference to calculate the coordinate offset of the target between adjacent frames to generate displacement vectors; calculate the ratio of the displacement vector to the corresponding time difference to obtain the real-time motion velocity of the target; standardize and encapsulate the motion trajectory, displacement vector, and real-time motion velocity to generate the behavioral features of the dynamic target.

7. The method according to claim 6, characterized in that, Step S600 specifically includes: Step S610: Real-time acquisition of structured semantic feature vectors and standardized business data with established mapping relationships; parsing and stripping the standardized business data to separate preset early warning rule data and business status data. Step S620: Based on a unified time base and a unified spatial identifier, extract multi-dimensional structured features from the structured semantic feature vector, and concatenate the multi-dimensional structured features with the corresponding business status data to generate fused feature data. Step S630: Input the fused feature data into the separated preset early warning rule data for logical judgment and generate analysis results; Step S640: When the generated analysis result meets the warning conditions, extract the globally unique tracking identifier and unified time base bound in the structured semantic feature vector that triggered the analysis result as the warning timestamp and generate the warning result; at the same time, trigger the first-level lightweight wake-up instruction and encapsulate and carry the globally unique tracking identifier and the warning timestamp into the first-level lightweight wake-up instruction.

8. The method according to claim 7, characterized in that, Step S630 specifically includes: Step S631: Extract structured semantic feature vectors containing target object category, motion trajectory and speed from the fused feature data, as well as business data containing train operation status, passenger flow status and equipment status; call the safety threshold matrix in the early warning rule data carried in the fused feature data. The safety threshold matrix is ​​a logical judgment set composed of speed limit, prohibited entry area identifier and density threshold corresponding to different dynamic target categories. Step S632: Compare the speed values ​​in the fused feature data with the speed upper limit values ​​in the safety threshold matrix, and calculate the speed difference between the two; if the speed difference is greater than zero, generate a first warning marker. Step S633: Perform a point-by-point intersection operation between the sequence of pixel coordinates corresponding to the motion trajectory in the fused feature data and the set of binary pixel coordinates corresponding to the prohibited entry area marker in the safety threshold matrix; if the intersection is not empty, it is determined that the motion trajectory and the prohibited entry area have spatial overlap, and a second warning marker is generated. Step S634: Obtain the train operation status from the business data and convert it into a status feature code; perform bit-by-bit matching of the status feature code with the preset linkage status logic rules in the safety threshold matrix; Step S635: The matching results of the first warning marker, the second warning marker, and the status feature code are weighted and summed to obtain the rule matching score; if the rule matching score reaches the preset logical threshold, the current scene is determined to meet the warning conditions, and the analysis result is generated.

9. The method according to claim 8, characterized in that, The specific process of step S700 includes: In step S710, the system retrieves the corresponding original image sequence from the circular buffer on the edge side based on the globally unique tracking identifier and warning timestamp carried in the first-level lightweight wake-up instruction; and locates and extracts the original foreground pixel stream from T seconds before the warning time to the current time. Step S720: Enter the adaptive ROI encoding mode, define the specific pixel area containing the warning target and its motion trajectory as the high-quality ROI area, and define the rest of the image except for the high-quality ROI area as the low-quality non-ROI area. Step S730: Perform non-uniform quantization encoding processing on the extracted pixel stream and assign quantization parameters QP to the high-quality ROI regions. roi To preserve the visual detail features of the target; assign quantization parameters QP to low-optimal non-ROI regions. non-roi QP reduces bitrate by drastically discarding redundant background information. non-roi >QP roi ; In step S740, the encoded lightweight foreground video stream is transmitted back to the server via the network. The server uses the spatial identifier and temporal reference in the transmitted stream to retrieve the corresponding static reference background image, executes pixel-level background restoration and fusion algorithms, and finally generates key event frames with complete environmental semantics.

10. The method according to claim 9, characterized in that, The specific process of step S800 includes: Step S810: Obtain the early warning results in real time and parse out the corresponding target type label, early warning time, and unified spatial identifier; Step S820: Based on the parsed target type label and unified spatial identifier, retrieve the preset action instruction mapping table; Step S830: Extract the handling action that matches the warning result from the action instruction mapping table, and combine it with the unified spatial identifier to locate the corresponding execution terminal, and generate a linkage handling instruction that includes the execution target, execution action and execution sequence; In step S840, the generated linkage handling command is output to the corresponding execution terminal, and the execution status is recorded synchronously.

Images