A parking lot abnormality alarm method and system based on space-time behavior modeling

Abstract

The application relates to the technical field of intelligent management and control of parking lots, in particular to a parking lot abnormality alarm method and system based on space-time behavior modeling, which comprises the following steps: S1. performing grid division on the global physical space of a parking lot, binding scene attributes, normal behavior time sequence rules and space interaction rules for the grid units, fusing space information and rule information to generate space-time semantic coding, and constructing a space-time semantic benchmark system. The application performs grid division on the global physical space of a parking lot and generates globally unique space-time semantic coding, constructs a space-time semantic benchmark system covering all scene behavior rules, converts target full-cycle behavior into a standardized space-time behavior semantic sequence, builds a space-time behavior cause-and-effect chain benchmark graph and a real-time behavior cause-and-effect chain graph, and incorporates the behavior time sequence continuity and space displacement correlation of personnel and vehicles in a parking lot into the core logic of abnormality determination.

Description

Technical Field

[0001] This invention relates to the field of intelligent parking management technology, specifically to a parking lot anomaly alarm method and system based on spatiotemporal behavior modeling. Background Technology

[0002] As frequently used enclosed or semi-enclosed public spaces in cities, parking lots experience dense pedestrian and vehicle traffic and complex environments. Timely identification and reliable alarm systems for abnormal events are crucial for ensuring the safe operation of parking lots and protecting the safety of people and property within them. Currently, the industry has widely adopted technologies such as sensor data collection and intelligent analysis to build parking lot safety monitoring and alarm systems that adapt to the safety management needs of parking lots of different sizes.

[0003] Among the existing published invention patents, publication number CN119964345A, the patent title is "A Parking Lot Emergency Early Warning Method and System Based on Multi-point Edge Computing". This solution collects multi-modal environmental and video data from edge nodes at multiple locations within the parking lot, constructs a three-dimensional environmental situation map, and combines multi-model collaborative analysis to complete the early warning and response to emergencies, providing a feasible implementation path for distributed safety monitoring of parking lots.

[0004] Existing technologies, in the process of identifying and judging abnormal events, only complete the event characterization based on the instantaneous state and spatial distribution of the collected data. They do not incorporate the temporal continuity and spatial displacement correlation of the behavior of people and vehicles in the parking lot into the core logic of abnormal judgment. This makes it impossible to identify and accurately alarm for gradual, non-sudden abnormal behaviors in advance. At the same time, it is easy to misjudge normal events that meet the instantaneous abnormality characteristics but have reasonable behavioral logic in terms of time and space as abnormal, and it is impossible to achieve a coordinated improvement in the advance and accuracy of abnormal alarms. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a parking lot anomaly alarm method and system based on spatiotemporal behavior modeling. This solves the problems of existing technologies that rely solely on instantaneous data to determine anomalies, are unable to proactively identify progressive non-sudden anomalies, have a high false alarm rate, and struggle to balance alarm proactiveness and identification accuracy.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a parking lot anomaly alarm method based on spatiotemporal behavior modeling, comprising the following steps: S1. Divide the entire physical space of the parking lot into grids, bind scene attributes, normal behavior timing rules and spatial interaction rules to grid units, integrate spatial information and rule information to generate spatiotemporal semantic codes, and construct a spatiotemporal semantic benchmark system; S2. Assign identity tags to targets within the parking lot, collect the target's behavior and spatial location data throughout the entire cycle, and match them to the spatiotemporal semantic encoding of the corresponding grid cell to generate a spatiotemporal behavior semantic sequence of the target throughout the entire cycle; S3. Extract the temporal dependencies, spatial transfer relationships, and causal logic relationships from the historical normal full-cycle spatiotemporal behavior semantic sequence, construct a baseline map of spatiotemporal behavior causal chains, and construct a real-time behavior causal chain map based on the target's real-time full-cycle spatiotemporal behavior semantic sequence. S4. Based on the real-time behavior causal chain graph and the spatiotemporal behavior causal chain benchmark graph, the subsequent trend of the target behavior is continuously deduced in multiple steps, the deviation of the deduced behavior from the normal behavior trend is calculated, an abnormal risk prediction value is generated, and potential abnormal behaviors are locked in. S5. Combining real-time full-cycle spatiotemporal behavioral semantic sequences with real-time behavioral causal chain graphs, progressive nested verification is performed from single-target temporal scale, single-grid spatial scale, and global interactive scale to identify and confirm abnormal behaviors and match risk levels. S6. Trigger alarm actions according to risk level, extract the misjudged and missed behaviors confirmed in the on-site handling results, optimize the temporal intensity weight of the directed edges between semantic nodes in the spatiotemporal behavior causal chain benchmark graph, and complete the incremental update of the benchmark graph.

[0007] Furthermore, in S1, the scene attributes bound to the grid cells include parking space area attributes, driving lane area attributes, pedestrian walkway area attributes, and non-public resting area attributes; The normal behavior timing rules include constraints on the maximum allowed dwell time in each area and constraints on the frequency of motion vector switching. The spatial interaction rules include the opening association rules between the target and the vehicle, and the crossing legality rules between the target and the grid boundary.

[0008] Furthermore, in S2, the specific process of generating the target's full-cycle spatiotemporal behavioral semantic sequence is as follows: The center point coordinates, motion direction and attitude features of the target are extracted by multi-sensor fusion perception technology. The timestamp information is spatiotemporally aligned with the spatiotemporal semantic encoding of the grid cell. The target's behavior data is filled into the corresponding semantic encoding container with discrete time steps to form a set of behavior vectors arranged linearly in time.

[0009] Furthermore, in S3, constructing a real-time behavioral causal chain graph includes: Each semantic code in the real-time full-cycle spatiotemporal behavioral semantic sequence is used as a graph node. Directed edges are established according to the order in which the behaviors occur. By calculating the transfer frequency and time interval between adjacent nodes, the directed edges are assigned a weight based on temporal intensity, thereby characterizing the logical order and time consumption distribution of the target transfer between different grid cells.

[0010] Furthermore, in S4, the multi-step continuous deduction of the subsequent trajectory of the target behavior includes: Starting from the latest behavior node in the real-time behavior causal chain graph, the high-probability subsequent transfer path of the corresponding node is retrieved in the spatiotemporal behavior causal chain baseline graph. The set of grid cells that the target may reach within the next N time steps is calculated, and the Euclidean distance between the actual observed behavior path and the predicted path is compared. If the Euclidean distance exceeds the set threshold, it is determined to be a deviation from the normal direction.

[0011] Furthermore, in S5, verification from a single target time-series scale refers to detecting abnormal changes in the target's acceleration and the frequency of its reciprocating motion within continuous time steps; Verification at a single-grid spatial scale refers to detecting the number of times a target appears in a restricted access area that does not match its identity attributes. Verification at the global interactive scale refers to detecting the distance contraction rate and the consistency of cooperative motion between the target and other identified targets.

[0012] Furthermore, in S5, the specific logic for identifying and confirming abnormal behavior and matching risk levels is as follows: If a single-scale verification is triggered, it is judged as low risk; If any two scale checks are triggered simultaneously, it is classified as medium risk. If all three scales trigger verification alarms and the abnormal risk prediction value exceeds the extreme value, it is judged as high risk, and the risk level information is passed through to the alarm response logic.

[0013] Furthermore, in S6, alarm actions triggered according to risk level include: Log and continuously track low-risk behaviors; For medium-risk behaviors, push early warning prompts to the management terminal and automatically retrieve the nearest monitoring footage; For high-risk behaviors, an audible and visual alarm is triggered on-site in the parking lot, and the real-time location of the target and the expected interception point are pushed to the mobile devices of security personnel.

[0014] Furthermore, in S6, the incremental update of the baseline map includes: Misjudged behaviors are treated as negative samples to reduce the judgment weight of the corresponding semantic node transition path in the spatiotemporal behavior causal chain baseline graph. The behavior sequences corresponding to the missed behaviors are marked as new abnormal patterns, the probability boundary distribution of normal behavior trends in the baseline graph is corrected, and the causal logic relationship matrix is ​​corrected in real time through incremental learning algorithms.

[0015] This invention also provides a parking lot anomaly alarm system based on spatiotemporal behavior modeling, comprising: The parking lot spatiotemporal semantic grid coding module is used to construct a spatiotemporal semantic benchmark system; The target full-cycle spatiotemporal behavior semantic serialization module is used to generate the target full-cycle spatiotemporal behavior semantic sequence. The dynamic construction module for spatiotemporal behavior causal chain graph is used to construct the baseline graph of spatiotemporal behavior causal chain and the real-time behavior causal chain graph. The multi-step behavioral evolution trend deduction module is used to predict the target behavior trend and generate abnormal risk prediction values; The multi-scale spatiotemporal nested verification module is used to perform progressive nested verification to confirm high-risk abnormal behavior and risk level; An anomaly classification pre-alarm module is used to trigger classified alarm actions and record the handling results; The causal chain map incremental self-calibration module is used to incrementally update the baseline map based on the treatment results.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a spatiotemporal semantic benchmark system covering all scenarios of behavior rules by dividing the entire physical space of a parking lot into a grid and generating globally unique spatiotemporal semantic codes. It transforms the target's full-cycle behavior into a standardized spatiotemporal behavioral semantic sequence, and builds a spatiotemporal behavioral causal chain benchmark map and a real-time behavioral causal chain map. It incorporates the temporal continuity and spatial displacement correlation of the behavior of people and vehicles in the parking lot into the core logic of anomaly judgment. This changes the limitation of existing technologies that only complete the qualitative analysis of events based on the instantaneous state and spatial distribution of collected data. It can perform pre-identification and risk prediction of progressive non-sudden abnormal behavior, while avoiding misjudgment of normal events that meet the instantaneous abnormality characteristics but have reasonable behavioral logic in terms of temporal sequence and space. Combined with multi-scale progressive nested verification, it completes the accurate classification and differentiated alarm handling of abnormal behavior. Then, based on the on-site handling results, it realizes incremental self-optimization of the benchmark map, continuously improving the accuracy and reliability of anomaly identification, and achieving a synergistic improvement in the pre-emptiveness and accuracy of anomaly alarms, providing stable and reliable technical support for the safe operation of parking lots. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the parking lot anomaly alarm method of the present invention; Figure 2 This is a flowchart of the multi-scale verification and risk level determination of abnormal behavior in this invention; Figure 3 This is a flowchart of the graded alarm handling and benchmark spectrum self-optimization process of the present invention; Figure 4 This is a diagram of the parking lot anomaly alarm system architecture of the present invention. Detailed Implementation

[0018] The technical solutions of 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.

[0019] This embodiment takes an underground enclosed parking lot in an urban commercial complex as the application object, and describes in detail a parking lot anomaly alarm method and system based on spatiotemporal behavior modeling. In this scenario, the parking lot has a total area of ​​approximately 20,000 square meters, including 800 parking spaces, a total driving lane length of 1,200 meters, and 6 pedestrian passages. It is equipped with multimodal sensing devices such as video surveillance cameras, millimeter-wave radar, and entrance and exit access control sensors to achieve full coverage of the parking lot without blind spots.

[0020] Please see Figure 1 This invention provides a parking lot anomaly alarm method based on spatiotemporal behavior modeling, comprising the following steps: S1. Divide the entire physical space of the parking lot into grids, bind scene attributes, normal behavior timing rules and spatial interaction rules to grid units, integrate spatial information and rule information to generate spatiotemporal semantic codes, and construct a spatiotemporal semantic benchmark system; S2. Assign identity tags to targets within the parking lot, collect the target's behavior and spatial location data throughout the entire cycle, and match them to the spatiotemporal semantic encoding of the corresponding grid cell to generate a spatiotemporal behavior semantic sequence of the target throughout the entire cycle; S3. Extract the temporal dependencies, spatial transfer relationships, and causal logic relationships from the historical normal full-cycle spatiotemporal behavior semantic sequence, construct a baseline map of spatiotemporal behavior causal chains, and construct a real-time behavior causal chain map based on the target's real-time full-cycle spatiotemporal behavior semantic sequence. S4. Based on the real-time behavior causal chain graph and the spatiotemporal behavior causal chain benchmark graph, the subsequent trend of the target behavior is continuously deduced in multiple steps, the deviation of the deduced behavior from the normal behavior trend is calculated, an abnormal risk prediction value is generated, and potential abnormal behaviors are locked in. S5. Combining real-time full-cycle spatiotemporal behavioral semantic sequences with real-time behavioral causal chain graphs, progressive nested verification is performed from single-target temporal scale, single-grid spatial scale, and global interactive scale to identify and confirm abnormal behaviors and match risk levels. S6. Trigger alarm actions according to risk level, extract the misjudged and missed behaviors confirmed in the on-site handling results, optimize the temporal intensity weight of the directed edges between semantic nodes in the spatiotemporal behavior causal chain benchmark graph, and complete the incremental update of the benchmark graph.

[0021] The above steps form a complete closed-loop process from static benchmark construction to dynamic behavioral semanticization, causal logic modeling, behavioral trend prediction, multi-dimensional verification and confirmation, hierarchical alarm handling, and benchmark self-optimization. This process incorporates the temporal continuity and spatial displacement correlation of the behavior of people and vehicles in the parking lot into the core logic of anomaly judgment. It changes the limitation of existing technologies that only complete the qualitative analysis of events based on the instantaneous state and spatial distribution of collected data. It can identify progressive, non-sudden abnormal behaviors in advance, while avoiding misjudgment of normal events that meet the characteristics of instantaneous anomalies but have reasonable behavioral logic in terms of temporal sequence and space. This achieves a synergistic improvement in the proactiveness and accuracy of anomaly alarms.

[0022] In a specific embodiment, in S1, the scene attributes bound to the grid cells include parking space area attributes, driving lane area attributes, pedestrian walkway area attributes, and non-public resting area attributes. The normal behavior timing rules include constraints on the maximum allowed dwell time in each area and constraints on the frequency of motion vector switching. The spatial interaction rules include the opening association rules between the target and the vehicle, and the crossing legality rules between the target and the grid boundary.

[0023] Specifically, this step uses an equidistant grid division method, dividing the entire physical space of the parking lot into square grid units with sides of 1 meter. Each grid unit has a unique spatial coordinate code, using a two-dimensional Cartesian coordinate system with the parking lot entrance as the origin, the X-axis horizontally, and the Y-axis vertically. The coordinate code format for each grid unit is as follows: ,in , These are the coordinates of the corresponding grid cell, in meters.

[0024] During the scene attribute binding process, the grid cells of the parking space area attribute correspond to the standard parking spaces, accessible parking spaces, and temporary passenger pick-up and drop-off spaces designated within the parking lot; the grid cells of the driving lane area attribute correspond to the vehicle driving lanes designated within the parking lot; the grid cells of the pedestrian passage area attribute correspond to the dedicated personnel passages, stairwells, and elevator lobby areas designated within the parking lot; and the grid cells of the non-public rest area attribute correspond to non-public open areas such as equipment rooms, fire control rooms, power distribution rooms, and management rooms within the parking lot.

[0025] The parameter values ​​of the normal behavior time sequence rule are obtained through the statistical analysis of the parking lot's historical normal operation data. The maximum stay duration constraint can be calculated by statistically analyzing the average stay duration of normal behavior in the corresponding area within the past 3 months, and taking 3 times the average value as the threshold. The motion vector switching frequency constraint refers to the number of times the target's movement direction is switched per unit time, which can be calculated by statistically analyzing the direction switching frequency of normal walking and normal driving behaviors, and taking the 95th percentile value as the upper limit threshold.

[0026] During the implementation of spatial interaction rules, the opening association rule between the target and the vehicle means that the opening interaction behavior between the target and the vehicle is deemed legal only when the spatial distance between the target and the corresponding vehicle is less than 0.5 meters and the vehicle door lock status changes from closed to open. The legality rule for the target to pass through the grid boundary means that the passing behavior between adjacent grid cells is deemed legal only when there is a physical connection path between grid cells and there are no physical isolation facilities such as walls, guardrails, or parking barriers. Passing across physically isolated grid cells is prohibited.

[0027] Spatiotemporal semantic coding adopts a fixed-length binary coding format with a total coding length of 32 bits. The first 8 bits are the scene attribute coding of the grid cell, the middle 12 bits are the normal behavior temporal rule coding, and the last 12 bits are the spatial interaction rule coding. The spatiotemporal semantic coding of each grid cell has global uniqueness and can directly map the spatial position and behavioral rule constraints of the corresponding grid, providing a unified benchmark system for the semantic matching of subsequent target behaviors.

[0028] In a specific embodiment, the process of generating the target's full-cycle spatiotemporal behavioral semantic sequence in S2 is as follows: The center point coordinates, motion direction and attitude features of the target are extracted by multi-sensor fusion perception technology. The timestamp information is spatiotemporally aligned with the spatiotemporal semantic encoding of the grid cell. The target's behavior data is filled into the corresponding semantic encoding container with discrete time steps to form a set of behavior vectors arranged linearly in time.

[0029] Specifically, the targets include three categories: people, motor vehicles, and non-motor vehicles. For targets entering the parking lot, multi-sensor fusion perception technology is used to detect and track targets. Each unique target is assigned a globally unique identification ID. The identification ID is encoded as a 16-bit string and contains three types of information: target type, entry time, and global serial number. This ensures that the target's identification is unique throughout the entire parking lot cycle and prevents tracking loss and ID switching.

[0030] The multi-sensor fusion perception technology combines the visual detection results of video surveillance cameras with the ranging and velocity measurement results of millimeter-wave radar. It uses the Kalman filter algorithm to achieve the optimal estimation of the target position and extracts the center point coordinates of the target. The center point coordinates are consistent with the grid coordinate system constructed in step S1 and can be directly matched to the corresponding grid cell.

[0031] The direction of motion is calculated by the change in the coordinates of the target center point within a continuous time step. The posture features for human targets include postures such as standing, walking, bending over, and running, and for vehicle targets include postures such as stationary, straight, reversing, and turning. The posture features are extracted and identified by a pre-trained deep learning classification model.

[0032] The spatiotemporal alignment of timestamp information and spatiotemporal semantic coding adopts a unified system clock, with a discrete time step set to 100ms. Each time step corresponds to a sampling period. Within each sampling period, the coordinates of the center point of the target are matched to the corresponding grid cell, and the spatiotemporal semantic coding of the grid cell is obtained. At the same time, the timestamp information of the current sampling period is bound to the spatiotemporal semantic coding to complete the spatiotemporal alignment.

[0033] The semantic encoding container is a pre-defined fixed-length data structure. Each semantic encoding container corresponds to a time step. The container stores six types of data: target identification ID, timestamp, spatiotemporal semantic code, center point coordinates, direction of movement, and posture features. It is a set of behavioral vectors arranged linearly in time. This is the sequence formed by arranging the semantic encoding containers corresponding to all time steps in the entire cycle of the target from entering the parking lot to leaving the parking lot in chronological order. In other words, it is the target's full-cycle spatiotemporal behavioral semantic sequence. This sequence transforms the target's continuous behavior into a discrete, quantifiable, and comparable set of semantic vectors, providing a data foundation for the subsequent construction of behavioral causal chains.

[0034] In a specific embodiment, S3, constructing a real-time behavioral causal chain graph includes: Each semantic code in the real-time full-cycle spatiotemporal behavioral semantic sequence is used as a graph node. Directed edges are established according to the order in which the behaviors occur. By calculating the transfer frequency and time interval between adjacent nodes, the directed edges are assigned a weight based on temporal intensity, thereby characterizing the logical order and time consumption distribution of the target transfer between different grid cells.

[0035] Specifically, the real-time behavior causal chain graph is defined as a directed weighted graph G=(V,E,W), where V is the set of graph nodes, each node corresponding to a spatiotemporal semantic code in the real-time full-cycle spatiotemporal behavior semantic sequence, E is the set of directed edges, and W is the set of weights for the directed edges. During the construction of the node set V, the spatiotemporal semantic code corresponding to the i-th time step in the target full-cycle spatiotemporal behavior semantic sequence is denoted as node V. , , N represents the total number of time steps within the entire target period, and the order of the nodes corresponds exactly to the order in which the actions occur. During the construction of the directed edge set E, nodes at adjacent time steps are selected according to the order in which the actions occur. and Establish directed edges between them , The directed edges point from the preceding node to the following node, representing the temporal sequence and causal logic of behaviors; that is, the preceding behavior is the cause of the following behavior, and the following behavior is the effect of the preceding behavior. The formulas involved in this step form a progressive calculation logic from single-dimensional feature quantification to two-dimensional feature fusion. First, the two single-dimensional indicators of spatial transfer rationality and time series compliance are calculated separately. Then, the final temporal intensity weight is obtained through weighted fusion. The application of each formula is based on the scenario adaptation of classic algorithms or the innovative fusion of this application, and the dimensions are kept highly consistent. At the same time, the calculation is implemented through actual parking lot operation data.

[0036] First, calculate the transition frequency between adjacent nodes. This refers to the sequence of normal historical behavior, starting from node... Transfer to node The number of times from node The proportion of all transfers originating from a given point is calculated using the following formula: This formula originates from the Markov chain state transition probability calculation method. This application treats each spatiotemporal semantic coding grid as a discrete state, and the movement of the target between grids as a state transition, adapting the classic formula to a specific scenario. At the dimensional level, both the numerator and denominator are transition counts, with a unified dimension of 1, resulting in a dimensionless probability value that meets the dimensional consistency requirement. The calculation is implemented based on six months of historical parking lot operation data, statistically analyzing data from the entrance lane node... The total number of departures was 12,000, of which transfers were to adjacent parking space nodes. The number of times was 9600. Substituting the data into the formula yields... This value indicates that the historical probability of this transfer behavior is 80%, suggesting that the spatial transfer is highly plausible.

[0037] Next, calculate the normalized value of the time interval between adjacent nodes. The calculation formula is: This formula originates from the classic ratio normalization method. Using the historical average transfer time as a benchmark, it eliminates the time dimension by using a ratio, quantifying the deviation of the actual transfer time from the normal benchmark. This represents a scenario-based adaptation of a classic algorithm. At the dimensional level, the time unit of both the numerator and denominator is seconds; after the ratio, the dimension is 1, unifying it with the dimension of the transfer frequency and laying the foundation for subsequent two-dimensional feature fusion. Continuing with the scenario of the above calculation example, in historical normal behavior... arrive The average transfer time is 2 seconds, and the actual transfer time of a certain real-time target is 2.1 seconds. Substituting the data into the formula, we can obtain... This value deviates very little from the baseline value of 1, indicating that the transfer time of the target conforms to the normal time pattern.

[0038] Finally, the directed edges are calculated. Temporal intensity weights The calculation formula is: α is a weighting coefficient, ranging from [0,1], used to balance the influence of transfer frequency and time interval on time series strength. The value of α can be determined through cross-validation of historical data. In this embodiment, α=0.6 is used to prioritize the weighting proportion of transfer frequency. The value range of the second term in the formula is (0,1]. When the actual time interval is consistent with the historical average time interval, the second term takes the maximum value of 1. As the time interval deviation increases, the value of the second term gradually decreases, and the final weight... The value range is (0,1], and the dimension is 1. The higher the weight value, the higher the degree of fit between the temporal logic of the transfer behavior and the normal behavior, and the stronger the causal rationality of the behavior. This formula is an innovative dual-feature linear weighted fusion formula in this application. It combines the two core causal dimensions of spatial transfer and time series, and corrects the penalty coefficient of time deviation through an inverse proportional function to achieve unified quantification of the rationality of spatiotemporal dual-dimensional behavior. The preceding transfer frequency formula and time interval normalization formula are the core inputs of this formula. The three form a clear progressive calculation logic and together constitute a complete calculation system of directed edge weights. At the dimension level, both inputs are dimensionless values. After weighted summation, the output dimension is still 1, which meets the dimensionless requirement of directed edge weights, and the dimension consistency is fully met. Continuing the above calculation scenario, take α=0.6, and calculate the... , Substituting into the formula, we get The weight value is in the higher range, indicating that the target is from arrive The transfer behavior has strong spatiotemporal causal rationality.

[0039] The construction of the spatiotemporal behavior causal chain benchmark graph is based on the semantic sequences of all normal full-cycle spatiotemporal behaviors within the past 6 months of parking lot history. The temporal dependencies, spatial transfer relationships, and causal logic relationships of all normal behaviors are extracted. Using the same graph structure and weight calculation method as the real-time behavior causal chain graph, a benchmark graph covering all normal behavior patterns of the parking lot is constructed. The benchmark graph includes the nodes, directed edges, and weight distributions corresponding to all normal behaviors in the parking lot, providing a benchmark reference for subsequent behavior inference and anomaly judgment.

[0040] In a specific embodiment, S4, the multi-step continuous deduction of the subsequent direction of the target behavior includes: Starting from the latest behavior node in the real-time behavior causal chain graph, the high-probability subsequent transfer path of the corresponding node is retrieved in the spatiotemporal behavior causal chain baseline graph. The set of grid cells that the target may reach within the next N time steps is calculated, and the Euclidean distance between the actual observed behavior path and the predicted path is compared. If the Euclidean distance exceeds the set threshold, it is determined to be a deviation from the normal direction.

[0041] Specifically, the step size N of the multi-step continuous extrapolation can be set according to the scene scale and behavioral characteristics of the parking lot. In this embodiment, for the underground parking lot of a commercial complex, N=50 is taken, corresponding to a behavioral extrapolation duration of 5 seconds, which can achieve accurate prediction of the short-term behavioral trend of the target. During the retrieval of high-probability transfer paths, the latest node in the real-time behavioral causal chain graph is used. Starting from this point, in the baseline graph of spatiotemporal behavioral causal chains, the search begins from... All directed edges are sorted from highest to lowest weight. The nodes corresponding to the first K directed edges are taken as the prediction nodes for the next step. In this embodiment, K=3 is taken to cover the main normal behavior transfer paths. This process is repeated for each prediction node, completing a multi-step continuous deduction for N time steps. This generates a set of grid cells that the target may reach within N time steps, as well as the corresponding prediction paths. The prediction paths are composed of the grid cell coordinates corresponding to the prediction nodes in each time step.

[0042] The actual observed behavior path refers to the grid cell coordinate sequence corresponding to the target behavior sequence actually collected within N time steps after the simulation is completed. The Euclidean distance between the predicted path and the actual path is calculated. For each time step, the Euclidean distance between the predicted coordinates and the actual coordinates is calculated, and then the average Euclidean distance over N time steps is taken as the final path deviation. The calculation formula is as follows: Where D is the average Euclidean distance along the path, in meters. , For the first Grid cell coordinates of the predicted path at each time step. , Here, represents the grid cell coordinates of the actual observed path at the nth time step, and N is the total number of time steps in the projection. This formula originates from the Euclidean distance calculation formula in Cartesian coordinates. This application, targeting multi-step time-series path scenarios, takes the arithmetic mean of the distances between points over N time steps to quantify the overall deviation of the entire path, representing a time-series scenario adaptation of the classic formula. In terms of dimensionality, the coordinate difference is measured in meters. Even after taking the square root, the dimension remains meters. The averaging of distances over multiple time steps maintains the meter dimension, fully meeting the dimensional requirements for path deviation, and achieving dimensional consistency.

[0043] The algorithm for this formula is as follows: First, a predicted coordinate sequence is generated. Then, the actual coordinate sequence is collected. Next, the Euclidean distance is calculated point by point. Then, the average value is calculated to obtain the path deviation D. Finally, D is compared with a set threshold to determine the abnormal path. The algorithm has a clear logical flow and can be implemented using actual coordinate data. A simplified calculation is performed using a step size of N=3. The predicted coordinate sequence is (1,2), (2,3), (3,4), and the actual coordinate sequence is (1.1,2.2), (2.2,3.1), (3.2,4.1). The single-step Euclidean distance is calculated point by point, and the result is approximately 0.224 meters. Substituting the data into the formula, we get D=(0.224+0.224+0.224) / 3=0.224 meters. If the set threshold obtained through historical normal behavior sequence statistics is 0.5 meters, and the calculated D value is less than the threshold, the target's behavior path is determined to be within the normal range. If the calculated D value exceeds the set threshold, the target's behavior is determined to be outside the normal range.

[0044] The threshold for Euclidean distance can be set by statistically analyzing the average Euclidean distance distribution between the predicted path and the actual path of historical normal behavior sequences, and taking the 95th percentile value as the threshold. When the calculated D exceeds the set threshold, it is determined that the target behavior deviates from the normal direction. At the same time, an abnormal risk prediction value is generated based on the deviation D. The abnormal risk prediction value is positively correlated with the deviation D. The higher the deviation, the larger the abnormal risk prediction value, and the higher the probability of abnormality of the target behavior, thereby achieving the pre-locking of potential abnormal behaviors.

[0045] In one specific embodiment, please refer to Figure 2 In S5, verification from a single target time scale refers to detecting abnormal changes in the target's acceleration and the frequency of its reciprocating motion within continuous time steps. Verification at a single-grid spatial scale refers to detecting the number of times a target appears in a restricted access area that does not match its identity attributes. Verification at the global interactive scale refers to detecting the distance contraction rate and the consistency of cooperative motion between the target and other identified targets.

[0046] Specifically, the execution logic of the progressive nested verification is as follows: first, single-target temporal scale verification is performed, then single-grid spatial scale verification is performed, and finally, global interactive scale verification is performed. The verification of the three scales is performed in sequence, covering three dimensions: target's own behavioral characteristics, spatial location compliance, and multi-target interactive behavior. This achieves comprehensive verification of abnormal behavior and avoids misjudgment and omission caused by single-dimensional verification.

[0047] During the single-target temporal scale verification process, the target's acceleration is calculated by the change in motion velocity within a continuous time step. Abnormal acceleration changes refer to the absolute value of acceleration exceeding a preset threshold, or the direction of acceleration switching frequently within a continuous time step, corresponding to abnormal behaviors such as sudden stops, rapid movements, and frequent changes in direction. The reciprocating motion frequency refers to the number of times the target moves back and forth within a fixed grid cell interval per unit time. The preset threshold for the reciprocating motion frequency is determined by the 95th percentile of the number of reciprocating motions per unit time in historical normal behavior. Exceeding this value is considered a temporal scale anomaly. When the reciprocating motion frequency exceeds the preset threshold, it is considered a temporal scale anomaly, corresponding to abnormal behaviors such as the target wandering or pacing within a fixed area. The preset threshold is determined by statistically analyzing the acceleration changes and reciprocating motion frequency distribution of historical normal behavior.

[0048] During the single-grid spatial scale verification process, the target's identity attributes include authorization attributes and type attributes. The authorization attribute is divided into authorized targets and unauthorized targets. Authorized targets include parking lot security personnel, maintenance personnel, cleaning personnel, etc., while unauthorized targets are ordinary social personnel and vehicles entering the parking lot. The type attribute is divided into three categories: personnel, motor vehicles, and non-motor vehicles. Restricted access areas refer to grid unit areas that do not match the target's identity attributes. For example, unauthorized personnel entering non-public rest areas, non-motor vehicles entering the core area of ​​motor vehicle lanes, and motor vehicles entering pedestrian walkways are all considered to be in restricted access areas that do not match their identity attributes. The number of times a target enters a restricted access area throughout the entire cycle is counted. When the number is greater than 0, it is judged as a spatial scale anomaly.

[0049] During the full-domain interactive scale verification process, the distance contraction rate between a target and other identified targets refers to the rate at which the spatial distance between two targets decreases per unit time. The preset threshold for the distance contraction rate is determined by the 95th percentile of the rate at which the distance between targets decreases in historical normal interactive behaviors. If the value exceeds this threshold, it is judged as an abnormal distance contraction, corresponding to the behavior of two targets rapidly approaching each other, which may pose abnormal risks such as tailing or conflict. The consistency of cooperative motion refers to the degree of overlap of the movement direction, speed, and trajectory of two or more targets within a continuous time step. If the degree of overlap exceeds the preset threshold and the duration exceeds the preset limit, it is judged as an abnormal cooperative motion consistency, corresponding to abnormal behaviors such as multiple people loitering together or multiple vehicles coordinating to occupy the road. The preset threshold is determined by statistically analyzing the distance changes and cooperative motion characteristics of historical normal interactive behaviors.

[0050] In one specific embodiment, please refer to Figure 2 In S5, the specific logic for identifying and confirming abnormal behavior and matching risk levels is as follows: If a single-scale verification is triggered, it is judged as low risk; If any two scale checks are triggered simultaneously, it is classified as medium risk. If all three scales trigger verification alarms and the abnormal risk prediction value exceeds the extreme value, it is judged as high risk, and the risk level information is passed through to the alarm response logic.

[0051] Specifically, the risk level matching adopts a progressive judgment logic, based on the verification results of three scales and the abnormal risk prediction value, to achieve accurate classification of abnormal behavior, avoid misjudgment of risk level caused by single-dimensional anomalies, and ensure priority response to high-risk abnormal behaviors.

[0052] Low-risk behavior is determined only when, among the three scales of single-target temporal scale, single-grid spatial scale, and global interactive scale, only one scale triggers an alarm, while the other two scales do not trigger an alarm, and the abnormal risk prediction value does not exceed the preset medium-risk threshold. Such behavior is usually instantaneous, non-malicious, has reasonable behavioral logic, and has no clear security risk.

[0053] The determination of medium-risk behavior is based on the fact that when the verification of any two of the three scales triggers an alarm simultaneously, and the abnormal risk prediction value exceeds the medium-risk threshold but does not exceed the high-risk extreme value, it is determined to be a medium-risk behavior. Such behavior has clear abnormal characteristics and potential security risks, which require management personnel to pay attention and verify in order to avoid the risk from escalating.

[0054] The determination of high-risk behavior is based on the following criteria: when all three scales of verification trigger alarms and the abnormal risk prediction value exceeds the preset high-risk extreme value, the behavior is determined to be high-risk. Such behavior has clear malice and security threat, and an emergency response needs to be triggered immediately to ensure the safety of people and property in the parking lot.

[0055] The extreme value of high-risk abnormality prediction is determined by statistically analyzing the distribution of risk prediction values ​​of historically confirmed high-risk abnormal events, and taking the 5th percentile value as the extreme value to ensure that all high-risk events that have occurred can be accurately identified.

[0056] In one specific embodiment, please refer to Figure 3 In S6, alarm actions triggered according to risk level include: Log and continuously track low-risk behaviors; For medium-risk behaviors, push early warning prompts to the management terminal and automatically retrieve the nearest monitoring footage; For high-risk behaviors, an audible and visual alarm is triggered on-site in the parking lot, and the real-time location of the target and the expected interception point are pushed to the mobile devices of security personnel.

[0057] Specifically, the execution of tiered alarm actions corresponds one-to-one with the risk level, enabling differentiated emergency responses. While ensuring safety, this avoids interference with management personnel caused by invalid alarms and improves the efficiency of parking lot security management.

[0058] During the handling of low-risk behaviors, the system automatically logs the full-cycle spatiotemporal semantic sequence of the behavior, the scale information that triggers the alarm, the target's identity and trajectory information, with a log storage period of no less than 90 days. At the same time, the system continuously tracks the target's subsequent behavior. If the subsequent behavior triggers a higher-level verification alarm, the risk level and response action are automatically upgraded. If the target leaves the parking lot normally, the tracking is automatically terminated, completing the log loop.

[0059] During the handling of medium-risk behaviors, the system immediately pushes an early warning to the fixed management terminal in the parking lot management room. The early warning includes the target's identity information, real-time location, abnormal characteristics at two scales that triggered the alarm, and behavioral trajectory information. At the same time, the system automatically retrieves real-time images from the three surveillance cameras closest to the target's current location and displays them in a split-screen format on the management terminal's screen, allowing managers to view the situation on-site in real time and verify the authenticity of the abnormal behavior. Meanwhile, the system continuously tracks the target's behavior. If the risk level escalates, it automatically switches to a high-risk response mode. If the target's abnormal behavior disappears, it reverts to a low-risk tracking mode. If the target leaves the parking lot normally, the early warning handling log is completed.

[0060] During the handling of high-risk behaviors, the system immediately triggers the audible and visual alarm devices in the corresponding area of ​​the parking lot, issuing an audible and visual warning to deter the perpetrator of the abnormal behavior. At the same time, the system pushes alarm information to the mobile devices of all on-duty security personnel. The information includes the target's real-time location, real-time video, abnormal behavior characteristics, expected movement path, and optimal interception point. The expected interception point is determined based on the target's real-time behavior trajectory and multi-step deduction results, providing precise guidance for security personnel's on-site handling. Simultaneously, the system automatically links with the access control devices at the parking lot entrances and exits, and can temporarily close the corresponding entrances and exits as needed to prevent the abnormal target from escaping, until the on-site handling is completed, and logs the entire alarm handling process.

[0061] During the handling of high-risk behaviors, the system immediately triggers the audible and visual alarm devices in the corresponding area of ​​the parking lot, issuing an audible and visual warning to deter the perpetrator of the abnormal behavior. At the same time, the system pushes alarm information to the mobile devices of all on-duty security personnel. The information includes the target's real-time location, real-time video, abnormal behavior characteristics, expected movement path, and optimal interception point. The expected interception point is determined based on the target's real-time behavior trajectory and multi-step deduction results, providing precise guidance for security personnel's on-site handling. Simultaneously, the system automatically links with the access control devices at the parking lot entrances and exits, and can temporarily close the corresponding entrances and exits as needed to prevent the abnormal target from escaping, until the on-site handling is completed, and logs the entire alarm handling process.

[0062] In one specific embodiment, please refer to Figure 3 In S6, the incremental update of the baseline map includes: Misjudged behaviors are treated as negative samples to reduce the judgment weight of the corresponding semantic node transition path in the spatiotemporal behavior causal chain baseline graph. The behavior sequences corresponding to the missed behaviors are marked as new abnormal patterns, the probability boundary distribution of normal behavior trends in the baseline graph is corrected, and the causal logic relationship matrix is ​​corrected in real time through incremental learning algorithms.

[0063] Specifically, the incremental updates of the baseline map are based on feedback from on-site handling results, enabling the system to self-learn and self-optimize, continuously improving the accuracy of anomaly identification and reducing the false positive and false negative rates. False positive behaviors refer to events that the system judges as abnormal but are verified as normal behavior through on-site handling. For such behaviors, their full-cycle spatiotemporal behavioral semantic sequence is used as a negative sample. The semantic nodes and transition paths corresponding to this behavioral sequence are extracted. In the spatiotemporal behavioral causal chain baseline map, the judgment weight of the corresponding transition path is reduced. The magnitude of the weight adjustment is positively correlated with the frequency of false positive behaviors. Simultaneously, the normal behavior probability boundary of the corresponding path is adjusted to prevent similar normal behaviors from being falsely judged as abnormal again. Missed detection refers to abnormal events that the system did not identify as abnormal, but were confirmed to pose a security risk upon on-site verification or post-event tracing. For such behaviors, the corresponding full-cycle spatiotemporal behavioral semantic sequence is labeled as a new abnormal pattern. The temporal features, spatial transition features, and interaction features of this abnormal pattern are extracted. The probability boundary distribution of the normal behavior trajectory in the baseline graph is corrected, and the behavioral features corresponding to the abnormal pattern are excluded from the normal behavior range. At the same time, through incremental learning algorithms, the causal logic relationship matrix of the baseline graph is corrected in real time, and the transition probability and weight distribution between nodes are updated so that the baseline graph can cover the newly added abnormal patterns and avoid the same type of abnormal event from being missed again.

[0064] The incremental learning algorithm employs online stochastic gradient descent, using a contrastive loss function as the optimization objective. The loss function formula is as follows: Where N is the number of samples, For the true labels of the samples, The model predicts probabilities. The initial learning rate is set to 0.01, and an exponential decay strategy is adopted with a decay coefficient of 0.9. The learning rate is updated every 100 iterations.

[0065] The causal logic relation matrix is ​​stored using a sparse matrix, with rows and columns corresponding to spatiotemporal semantic coding nodes. Matrix elements represent the temporal strength weights of directed edges between nodes, and non-zero elements retain only the weights of valid transition paths. The incremental learning algorithm employs an online stochastic gradient descent algorithm, using confirmed normal and abnormal behaviors as training samples and the causal logic relation matrix of the baseline graph as the optimization objective. This allows for real-time updates of model parameters without interrupting normal system operation, achieving online incremental updates of the baseline graph and ensuring that the system's anomaly recognition capability continuously adapts to the ever-changing behavioral patterns and scenario characteristics during parking lot operation.

[0066] Please see Figure 4 The present invention also provides a parking lot anomaly alarm system based on spatiotemporal behavior modeling, comprising: The parking lot spatiotemporal semantic grid coding module is used to construct a spatiotemporal semantic benchmark system; The target full-cycle spatiotemporal behavior semantic serialization module is used to generate the target full-cycle spatiotemporal behavior semantic sequence. The dynamic construction module for spatiotemporal behavior causal chain graph is used to construct the baseline graph of spatiotemporal behavior causal chain and the real-time behavior causal chain graph. The multi-step behavioral evolution trend deduction module is used to predict the target behavior trend and generate abnormal risk prediction values; The multi-scale spatiotemporal nested verification module is used to perform progressive nested verification to confirm high-risk abnormal behavior and risk level; An anomaly classification pre-alarm module is used to trigger classified alarm actions and record the handling results; The causal chain map incremental self-calibration module is used to incrementally update the baseline map based on the treatment results.

[0067] Specifically, the system adopts a deployment architecture that combines edge computing and cloud computing. The parking lot spatiotemporal semantic grid encoding module, the target full-cycle spatiotemporal behavior semantic serialization module, the spatiotemporal behavior causal chain graph dynamic construction module, the behavior evolution trend multi-step inference module, and the multi-scale spatiotemporal nested verification module are deployed on the edge computing nodes at the parking lot site to achieve real-time processing of on-site data and local identification of abnormal behavior, reduce data transmission latency, and ensure the real-time performance of alarm response. The anomaly classification pre-alarm module and the causal chain graph incremental self-calibration module are deployed on the cloud management platform to achieve centralized management of multiple parking lots, unified scheduling of alarm information, and global optimization and update of the baseline graph.

[0068] In summary, this invention constructs a spatiotemporal semantic benchmark system covering all scenarios of behavior rules by dividing the entire physical space of a parking lot into a grid and generating globally unique spatiotemporal semantic codes. It transforms the target's full-cycle behavior into a standardized spatiotemporal behavioral semantic sequence, and builds a spatiotemporal behavioral causal chain benchmark map and a real-time behavioral causal chain map. It incorporates the temporal continuity and spatial displacement correlation of the behavior of people and vehicles within the parking lot into the core logic of anomaly judgment. This overcomes the limitations of existing technologies that only rely on the instantaneous state and spatial distribution of collected data to qualitatively classify events. It can perform pre-identification and risk prediction of progressive, non-sudden abnormal behaviors, while avoiding misjudgment of normal events that meet the characteristics of instantaneous anomalies but have reasonable behavioral logic in terms of temporal sequence and space. Combined with multi-scale progressive nested verification, it achieves accurate classification and differentiated alarm handling of abnormal behaviors. Furthermore, based on the on-site handling results, it achieves incremental self-optimization of the benchmark map, continuously improving the accuracy and reliability of anomaly identification. This achieves a synergistic improvement in the pre-emptiveness and accuracy of anomaly alarms, providing stable and reliable technical support for the safe operation of parking lots.

[0069] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0070] 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 parking lot anomaly alarm method based on spatiotemporal behavior modeling, characterized in that, Includes the following steps: S1. Divide the entire physical space of the parking lot into grids, bind scene attributes, normal behavior timing rules and spatial interaction rules to grid units, integrate spatial information and rule information to generate spatiotemporal semantic codes, and construct a spatiotemporal semantic benchmark system; S2. Assign identity tags to targets within the parking lot, collect the target's behavior and spatial location data throughout the entire cycle, and match them to the spatiotemporal semantic encoding of the corresponding grid cell to generate a spatiotemporal behavior semantic sequence of the target throughout the entire cycle; S3. Extract the temporal dependencies, spatial transfer relationships, and causal logic relationships from the historical normal full-cycle spatiotemporal behavior semantic sequence, construct a baseline map of spatiotemporal behavior causal chains, and construct a real-time behavior causal chain map based on the target's real-time full-cycle spatiotemporal behavior semantic sequence. The construction of the real-time behavioral causal chain graph includes: Each semantic code in the real-time full-cycle spatiotemporal behavioral semantic sequence is used as a graph node. Directed edges are established according to the order in which the behaviors occur. By calculating the transfer frequency and time interval between adjacent nodes, the directed edges are assigned a weight based on temporal intensity, thereby characterizing the logical order and time consumption distribution of the target transfer between different grid cells. S4. Based on the real-time behavior causal chain graph and the spatiotemporal behavior causal chain benchmark graph, the subsequent trend of the target behavior is continuously deduced in multiple steps, the deviation of the deduced behavior from the normal behavior trend is calculated, an abnormal risk prediction value is generated, and potential abnormal behaviors are locked in. The multi-step continuous deduction of the subsequent trajectory of the target behavior includes: Starting from the latest behavior node in the real-time behavior causal chain graph, the high-probability subsequent transfer path of the corresponding node is retrieved in the spatiotemporal behavior causal chain baseline graph. The set of grid cells that the target may reach within the next N time steps is calculated, and the Euclidean distance between the actual observed behavior path and the predicted path is compared. If the Euclidean distance exceeds the set threshold, it is determined to be a deviation from the normal direction. S5. Combining real-time full-cycle spatiotemporal behavioral semantic sequences with real-time behavioral causal chain graphs, progressive nested verification is performed from single-target temporal scale, single-grid spatial scale, and global interactive scale to identify and confirm abnormal behaviors and match risk levels. Among them, verification from a single target time scale refers to detecting abnormal changes in the target's acceleration and the frequency of its reciprocating motion within continuous time steps; Verification at a single-grid spatial scale refers to detecting the number of times a target appears in a restricted access area that does not match its identity attributes. Verification at the global interactive scale refers to detecting the distance contraction rate and the consistency of cooperative motion between the target and other identified targets; S6. Trigger alarm actions according to risk level, extract the misjudged and missed behaviors confirmed in the on-site handling results, optimize the temporal intensity weight of the directed edges between semantic nodes in the spatiotemporal behavior causal chain benchmark graph, and complete the incremental update of the benchmark graph.

2. The parking lot anomaly alarm method based on spatiotemporal behavior modeling according to claim 1, characterized in that, In S1, the scene attributes bound to the grid cells include parking space area attributes, driving lane area attributes, pedestrian walkway area attributes, and non-public rest area attributes. The normal behavior timing rules include constraints on the maximum allowed dwell time in each area and constraints on the frequency of motion vector switching. The spatial interaction rules include the opening association rules between the target and the vehicle, and the crossing legality rules between the target and the grid boundary.

3. The parking lot anomaly alarm method based on spatiotemporal behavior modeling according to claim 1, characterized in that, In S2, the specific process of generating the target's full-cycle spatiotemporal behavioral semantic sequence is as follows: The center point coordinates, motion direction and attitude features of the target are extracted by multi-sensor fusion perception technology. The timestamp information is spatiotemporally aligned with the spatiotemporal semantic encoding of the grid cell. The target's behavior data is filled into the corresponding semantic encoding container with discrete time steps to form a set of behavior vectors arranged linearly in time.

4. The parking lot anomaly alarm method based on spatiotemporal behavior modeling according to claim 1, characterized in that, In S5, the specific logic for identifying and confirming abnormal behavior and matching risk levels is as follows: If a single-scale verification is triggered, it is judged as low risk; If any two scale checks are triggered simultaneously, the risk level is determined to be medium. If all three scales trigger verification alarms and the abnormal risk prediction value exceeds the extreme value, it is judged as high risk, and the risk level information is passed through to the alarm response logic.

5. The parking lot anomaly alarm method based on spatiotemporal behavior modeling according to claim 1, characterized in that, In S6, alarm actions triggered according to risk level include: Log and continuously track low-risk behaviors; For medium-risk behaviors, push early warning prompts to the management terminal and automatically retrieve the nearest monitoring footage; For high-risk behaviors, an audible and visual alarm is triggered on-site in the parking lot, and the real-time location of the target and the expected interception point are pushed to the mobile devices of security personnel.

6. The parking lot anomaly alarm method based on spatiotemporal behavior modeling according to claim 1, characterized in that, In S6, the incremental update of the baseline map includes: Misjudged behaviors are treated as negative samples to reduce the judgment weight of the corresponding semantic node transition path in the spatiotemporal behavior causal chain baseline graph. The behavior sequences corresponding to the missed behaviors are marked as new abnormal patterns, the probability boundary distribution of normal behavior trends in the baseline graph is corrected, and the causal logic relationship matrix is ​​corrected in real time through incremental learning algorithms.

7. A parking lot anomaly alarm system based on spatiotemporal behavior modeling, used to implement the parking lot anomaly alarm method based on spatiotemporal behavior modeling as described in any one of claims 1-6, characterized in that, include: The parking lot spatiotemporal semantic grid coding module is used to construct a spatiotemporal semantic benchmark system; The target full-cycle spatiotemporal behavior semantic serialization module is used to generate the target full-cycle spatiotemporal behavior semantic sequence. The dynamic construction module for spatiotemporal behavior causal chain graph is used to construct the baseline graph of spatiotemporal behavior causal chain and the real-time behavior causal chain graph. The multi-step behavioral evolution trend deduction module is used to predict the target behavior trend and generate abnormal risk prediction values; The multi-scale spatiotemporal nested verification module is used to perform progressive nested verification to confirm high-risk abnormal behavior and risk level; An anomaly classification pre-alarm module is used to trigger classified alarm actions and record the handling results; The causal chain map incremental self-calibration module is used to incrementally update the baseline map based on the treatment results.

Images