High-rise residential personnel abnormal behavior early warning method and system based on behavior baseline

Abstract

The application discloses a high-rise residential personnel abnormal behavior early warning method and system based on a behavior baseline, and relates to the technical field of intelligent security and protection. The method comprises the following steps: positioning a target object of a high-rise residence to perform real-time multi-source perception fusion, and generating a fusion behavior data stream; performing spatio-temporal context behavior analysis on the fusion behavior data stream, and extracting a structured behavior event sequence; performing group behavior analysis based on the structured behavior event sequence, setting an initial individual behavior baseline, performing incremental learning update, and generating an individual behavior baseline; performing multi-dimensional deviation calculation on the structured behavior event sequence and the individual behavior baseline, and obtaining a multi-dimensional abnormal deviation degree; triggering a multi-level early warning condition, and generating multi-level abnormal behavior early warning information to early warn a high-rise residential personnel abnormal behavior. The technical problems of low identification accuracy and high false alarm and missed alarm rates of a high-rise residential personnel abnormal behavior in the prior art are solved, and the technical effects of accurate identification of an abnormal behavior and reduction of a false alarm rate are achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent security technology, specifically to a method and system for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines. Background Technology

[0002] In densely populated residential settings such as high-rise buildings, human activity is highly random and complex, involving various spatial interactions such as those occurring in elevators, public corridors, and entrances / exits. Traditional security systems primarily rely on video surveillance or single sensors for behavior recognition, typically judging human behavior based on fixed rules or static thresholds. However, such methods struggle to effectively characterize the behavioral differences between individuals in complex scenarios, resulting in limited accuracy in anomaly detection. Furthermore, under the influence of multiple interference factors, the system is prone to misjudgments and missed detections, thereby reducing the overall reliability and practicality of early warning systems. Summary of the Invention

[0003] This application provides a method and system for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines, which solves the technical problems of low accuracy and high false alarm and missed alarm rates in the identification of abnormal behavior of people in high-rise residential buildings in the prior art.

[0004] The first aspect of this application provides a method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines, the method comprising: The system locates target objects in high-rise residential buildings and performs real-time multi-source perception fusion to generate a fused behavioral data stream. It then performs spatiotemporal context behavior analysis on the fused behavioral data stream to extract a structured behavioral event sequence of the target objects. Based on the structured behavioral event sequence, it performs group behavior analysis, sets an initial individual behavioral baseline for the target objects, and incrementally learns and updates the data according to this baseline to generate individual behavioral baselines. It then calculates a multi-dimensional deviation between the structured behavioral event sequence and the individual behavioral baseline to obtain a multi-dimensional anomaly deviation, which includes an anomaly score. Based on the anomaly score, it triggers multi-level early warning conditions to generate multi-level abnormal behavior early warning information to warn of abnormal behavior among residents in high-rise residential buildings.

[0005] A second aspect of this application provides an early warning system for abnormal behavior of residents in high-rise residential buildings based on behavioral baselines, the system comprising: The system comprises the following modules: a data fusion module, a behavior analysis module, and a behavior analysis module. The data fusion module performs real-time multi-source perception fusion on target objects in high-rise residential buildings, generating a fused behavior data stream. The behavior analysis module performs spatiotemporal context behavior analysis on the fused behavior data stream, extracting a structured behavior event sequence from the target objects. The behavior analysis module performs group behavior analysis based on the structured behavior event sequence, sets an initial individual behavior baseline for the target objects, and performs incremental learning and updates according to the initial individual behavior baseline to generate an individual behavior baseline. The deviation calculation module performs multi-dimensional deviation calculation between the structured behavior event sequence and the individual behavior baseline to obtain a multi-dimensional abnormal deviation degree, which includes an abnormal score. The early warning module triggers multi-level early warning conditions based on the abnormal score, generating multi-level abnormal behavior early warning information to warn of abnormal behavior among residents in high-rise residential buildings.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, the target objects in high-rise residential buildings are located and subjected to real-time multi-source perception fusion to generate a fused behavioral data stream. Next, the fused behavioral data stream is analyzed for spatiotemporal context to extract structured behavioral event sequences of the target objects. Further, based on the structured behavioral event sequences, group behavior analysis is performed to establish an initial individual behavioral baseline for the target objects. Incremental learning and updates are then performed according to this initial baseline to generate individual behavioral baselines. Then, multi-dimensional deviation calculations are performed between the structured behavioral event sequences and the individual behavioral baselines to obtain a multi-dimensional anomaly deviation, which includes an anomaly score. Finally, multi-level early warning conditions are triggered based on the anomaly score to generate multi-level abnormal behavior early warning information for alerting residents in high-rise residential buildings to report abnormal behaviors. This approach solves the technical problems of low accuracy and high false alarm / missed alarm rates in existing technologies for identifying abnormal behaviors in high-rise residential buildings, achieving accurate identification of abnormal behaviors and reducing the false alarm rate. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines provided in this application embodiment; Figure 2 This application provides a multi-level early warning triggering diagram for its embodiments. Figure 3A schematic diagram of the structure of a high-rise residential building occupant abnormal behavior early warning system based on behavioral baselines provided in this application embodiment.

[0009] Figure labeling: Data fusion module 11, Behavior analysis module 12, Behavior analysis module 13, Deviation calculation module 14, Early warning module 15. Detailed Implementation

[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0011] Example 1, as Figure 1 As shown, this application provides a method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines, wherein the method includes: The system locates target objects in high-rise residential buildings and performs real-time multi-source perception fusion to generate a fused behavioral data stream.

[0012] Furthermore, the method for locating target objects in high-rise residential buildings and performing real-time multi-source perception fusion to generate a fused behavioral data stream includes: Multi-view synchronous acquisition of public areas in high-rise residential buildings generates multi-view video streams, simultaneously acquiring access control event streams, floor infrared trigger data, and point cloud data of the public areas. Multi-target re-identification is performed on the multi-view video streams to obtain cross-camera trajectory segments. Target clustering and tracking are performed on the point cloud data to obtain radar target trajectories. The access control event streams and cross-camera trajectory segments are timestamped to generate target object identification identifiers. Continuous trajectory identification is performed by traversing the radar target trajectories to determine missing trajectory data segments. Data interpolation is performed based on the missing trajectory data segments to complete the continuous full trajectory. Spatial constraint registration is performed between the continuous full trajectory and the floor infrared trigger data to generate a multimodal time series, which is then used as the fused behavioral data stream.

[0013] Multi-view video acquisition equipment, access control and recognition equipment, floor infrared trigger sensors, and 3D point cloud acquisition equipment are deployed in the public areas, elevator lobbies, corridors, and entrances of high-rise residential buildings to achieve multimodal real-time perception of target objects. The multi-view video acquisition equipment synchronously acquires images of each monitoring area at a preset frame rate and uses a time synchronization module to uniformly calibrate the acquisition clocks of different cameras to obtain a unified time reference. The access control and recognition equipment collects card swipe records, identity verification results, and door opening events of target objects in real time and generates a timestamped access control event stream. The floor infrared trigger sensors are used to detect the movement of target objects on floors or in corridors. A 3D point cloud acquisition device, triggered by a signal and outputting a corresponding spatial location marker, performs spatial scanning of the target object and its surrounding environment, generating a point cloud data sequence with depth information. A multi-view video acquisition device continuously acquires images of the target object and performs identity matching and cross-camera association of the target object from different camera perspectives based on a preset multi-target re-identification algorithm. This algorithm extracts target appearance feature vectors and motion trajectory features and performs joint encoding, calculating similarity within a unified feature space to achieve cross-camera target association, thereby obtaining cross-camera trajectory segments. The time continuity of these cross-camera trajectory segments is then checked to ensure... The system ensures the consistency of the trajectory timing; simultaneously, it collects card swipe records, identity recognition records, and door opening event streams output by access control devices, and performs time deviation alignment processing on different data sources based on a unified timestamp mechanism. It then performs time synchronization matching between the access control event stream and cross-camera trajectory segments. When the time difference meets a preset threshold, it establishes an identity association to determine the unique identity of the target object. Furthermore, it performs spatial clustering and target tracking processing on the point cloud data, using a clustering method based on distance thresholds and density distribution to extract the spatial point set of the target object. It then generates a spatial motion trajectory based on continuous frame association tracking, simultaneously identifying missing and occluded segments in the trajectory, and processing the discontinuous segments... The segment employs an interpolation method based on motion continuity constraints to complete the trajectory and obtain a continuous full trajectory. Subsequently, the completed spatial motion trajectory is mapped and constrained with the floor infrared trigger data through spatial coordinate system mapping. By establishing a unified floor spatial reference coordinate system, the spatial alignment of point cloud coordinates and infrared trigger positions is achieved to correct multi-source perception bias. Finally, multi-view video data, access control event data, infrared trigger data, and point cloud trajectory data are fused and encoded according to a unified time axis. A multi-modal time series containing time information, spatial information, and identity association information is constructed by timestamp sorting and multi-modal feature splicing, and this is output as a fused behavioral data stream.

[0014] The spatiotemporal context behavior parsing of the fused behavior data stream is performed to extract the structured behavior event sequence of the target object.

[0015] Furthermore, the spatiotemporal context behavior parsing of the fused behavioral data stream is performed to extract the structured behavioral event sequence of the target object. The method includes: A spatial topology of a high-rise residential building is constructed. Based on this spatial topology and the fused behavioral data stream, a spatiotemporal location map of the target object is constructed. A temporal graph attention network is used to extract features from the spatiotemporal location map, generating a sequence of behavioral node embedding vectors. Behavior recognition is performed based on the sequence of behavioral node embedding vectors to identify and determine the atomic behavior categories of the target object, obtaining a sequence of atomic behavior categories. A temporal segmentation network is used to detect event boundaries in the sequence of atomic behavior categories, determining multiple behavioral event boundary data, which include event start time and event end time. Feature extraction is performed according to the event start time, the event end time, and the sequence of atomic behavior categories to obtain spatiotemporal attribute features and object association features. The structured behavioral event sequence is constructed based on the spatiotemporal attribute features and the object association features.

[0016] Specifically, the building floor plan and floor structure information of the high-rise residential building are read, and elevators, stairs, corridors, public halls, entrances and exits and various functional areas are defined as graph nodes. The passability relationship between adjacent areas is defined as graph edges. Each node is assigned a unique node number and spatial coordinates to construct the spatial topology of the high-rise residential building. Then, the behavior data stream is traversed and fused, and the spatial location coordinates of the target object are read in the order of timestamps. The location coordinates of the target object at each sampling time are mapped to the corresponding spatial topology node, and directed connection relationships between nodes are established in the order of continuous time to form the spatiotemporal location graph of the target object. Each node records the node number, entry time, exit time, dwell time and movement speed, and each directed edge records the movement direction and movement time of the target object between adjacent nodes.

[0017] The spatiotemporal location map is input into a temporal graph attention network. The temporal graph attention network first uses a graph attention layer to calculate the attention weights between the current node and its neighboring nodes, then performs a weighted summation of the features of neighboring nodes according to these attention weights to obtain the spatial association features of the current node. Subsequently, a gated recurrent unit is used to recursively update the node features across multiple consecutive time slices in chronological order, obtaining a sequence of behavioral node embedding vectors containing both temporal and spatial associations. This sequence of behavioral node embedding vectors is then input into a fully connected classification layer, and a Softmax classifier is used to output the corresponding behavioral categories, determining the atomic behavior categories of the target object within each consecutive time slice. These atomic behaviors include dwelling, walking, going up and down stairs, entering an elevator, leaving an elevator, entering a room, and leaving a room. The sequence of atomic behavior categories is then input into a temporal segmentation network, where a one-dimensional convolutional network extracts local temporal features, and a bidirectional long short-term memory network extracts long-term temporal dependencies. Finally, the change points in the classification results are used to detect the start and end positions of events, determining the boundary data for each behavioral event. The boundary data for each behavioral event includes at least the event start time and the event end time.

[0018] Based on the boundaries of each behavioral event, behavioral data within the corresponding time period is extracted. The floor where the event occurred, the functional area, the duration of stay, the distance traveled, the average speed of movement, and the time of the event are statistically analyzed as spatiotemporal attribute features. At the same time, the structured trajectories of other target objects within the same time window are traversed. When the spatial distance between two target objects is less than a preset distance threshold and the duration exceeds a preset contact time threshold, the identity of the associated object, the duration of the association, and the type of the associated behavior are recorded as object association features. Finally, the event is uniformly encapsulated according to the event number, event type, event start time, event end time, spatiotemporal attribute features, and object association features to generate a structured behavioral event sequence.

[0019] The temporal graph attention network is implemented using a structure of "graph attention layer - gated recurrent unit - GRU - fully connected feature fusion layer". Specifically, it includes: taking the spatiotemporal location graph as the network input, where each node corresponds to the state information of the target object at a time sampling point. The state information includes at least spatial coordinates, floor number, functional area number, timestamp, movement speed and movement direction. Each edge represents the spatial transfer relationship and temporal continuity relationship between adjacent time slices. Specifically, firstly, the node state information is input into the graph attention layer. A linear transformation is performed on each node to obtain a node feature vector, and the similarity coefficient between the current node and its neighboring nodes is calculated. Based on the similarity coefficient, the attention weight is obtained by normalization using the Softmax function. Then, the attention weight is used to perform a weighted summation of the neighboring node features to obtain a node representation vector that integrates the neighborhood spatial relationships. Subsequently, the node representation vectors arranged in chronological order are input into a gated recurrent unit (GRU). The historical state and the current input are recursively updated through update and reset gates to obtain a hidden state vector containing time dependencies. The hidden state vectors are then input into a fully connected feature fusion layer for nonlinear mapping and dimensionality compression to generate a fixed-dimensional behavior node embedding vector. Finally, the behavior node embedding vectors are output in chronological order to form a behavior node embedding vector sequence for subsequent atomic behavior category recognition processing.

[0020] Furthermore, based on the spatial topology and the fused behavioral data stream, a spatiotemporal location map of the target object is constructed, the method includes: Based on the spatial topology of the high-rise residential building, functional zones are defined, and multiple functional areas are delineated. The fusion behavior data stream is traversed for location matching and identification to determine multiple location points. Timestamps are extracted from the fusion behavior data stream, and the multiple location points are mapped to the multiple functional areas according to the timestamps to construct region-time tuple nodes. Based on the region-time tuple nodes, the functional area change analysis of the target object between adjacent time slices is performed, and multiple directed edges of nodes are established. The multiple directed edges of nodes are added to the spatiotemporal location map.

[0021] Preferably, based on the building plan data, floor connectivity, and public area functional attributes of the high-rise residential building, the spatial topology is functionally partitioned, delineating multiple functional areas such as elevator lobbies, corridors, stairwells, access control entrances / exits, public platforms, and target floor access areas, and assigning a unique area identifier to each functional area; subsequently, the fused behavioral data stream is traversed to extract the location coordinates, floor number, and perception source identifier of the target object at each timestamp, and the location coordinates are matched with the spatial boundaries of the preset functional areas to determine multiple location points of the target object at different time points; based on the timestamp information in the fused behavioral data stream, the multiple location points are mapped to the corresponding functional areas in chronological order, constructing a system consisting of area identifiers, timestamps, and target objects. The system consists of a region-time tuple node composed of identity identifiers and location confidence scores. Further, the region-time tuple nodes are sorted according to a preset time slice interval. The system compares whether the functional area to which the target object belongs changes within adjacent time slices. If the functional areas corresponding to adjacent time slices are different, a directed edge representing the regional migration relationship is established between the region-time tuple node corresponding to the previous time slice and the region-time tuple node corresponding to the next time slice. If the functional areas corresponding to adjacent time slices are the same, a directed edge representing the regional residence relationship is established. Finally, the region-time tuple nodes and the multiple directed edges are written together into the graph structure data to generate a spatiotemporal location map representing the target object's movement, residence, and regional switching relationships within the functional areas of a high-rise residential building over time.

[0022] Group behavior analysis is performed based on the structured behavioral event sequence. An initial individual behavior baseline of the target object is set, and incremental learning and updating are performed according to the initial individual behavior baseline of the target object to generate an individual behavior baseline.

[0023] Furthermore, based on the structured behavioral event sequence, group behavior analysis is performed to establish an initial individual behavioral baseline for the target group. The method includes: A group behavior baseline is constructed, and the attribute features of the target object are extracted and matched with the group behavior baseline to determine a group behavior template, which includes a set of frequent sequence patterns. Attention distribution encoding is performed based on the set of frequent sequence patterns to construct a group behavior prototype vector. An individual behavior space is constructed by mapping the group behavior prototype vector to the individual behavior space using the structured behavior event sequence as a constraint, and calculating the initial event probability distribution data, the initial event duration distribution data, and the initial event transition probability matrix. The initial event probability distribution data, the initial event duration distribution data, and the initial event transition probability matrix are integrated to construct the initial individual behavior baseline.

[0024] First, a sequence of structured behavioral events from the past was collected for high-rise residential buildings. Then, resident attribute vectors were constructed according to age group, family size, housing type, and daily routine. Euclidean distance was used to calculate the distance between the target object attribute vector and the attribute centers of each group. in, Represents the target object attribute vector. 3D eigenvalues Let m represent the feature value of the j-th dimension corresponding to the i-th group center, and m represent the dimension of the attribute feature. The group with the smallest distance is determined as the behavioral group to which the target object belongs, and the frequent sequence pattern set in the corresponding group behavior template is called. Then, each frequent sequence pattern is encoded into a behavioral feature vector, and the query vector Q, key vector K, and value vector V are calculated respectively. The importance of each frequent sequence pattern is calculated using scaled dot product attention. Where d represents the vector dimension and N represents the number of frequent sequence patterns. This represents the attention weight corresponding to the i-th frequent sequence pattern; further, the group behavior prototype vector is calculated based on the attention weight: Where P represents the group behavior prototype vector. Let i represent the value vector corresponding to the i-th frequent sequence pattern; then, construct an individual behavior space, using the structured behavioral event sequence of the target object for the most recent consecutive preset number of days as constraints, map the group behavior prototype vector to the individual behavior space, and count the occurrence frequency of each event type. and the total number of all events Calculate the initial event probability distribution data: in, Indicates event type The probability of occurrence; further statistical analysis of the duration sample set of each event, and calculation of the mean duration of the events: and standard deviation: in, Let represent the duration of the event type i when it occurs for the kth time, and construct the initial event duration distribution data accordingly; then count the number of times any event type i transitions to event type j. Calculate the event transition probability: Where M represents the total number of event types, The probability of transitioning from event type i to event type j is represented, and an initial event transition probability matrix is ​​constructed accordingly. Finally, the initial event probability distribution data, the initial event duration distribution data, and the initial event transition probability matrix are uniformly encapsulated to form the initial individual behavior baseline of the target object.

[0025] Furthermore, the process of constructing a baseline for group behavior includes the following methods: A collection of historical behavioral event sequences of residents in high-rise residential buildings is gathered. These sequences are then grouped according to resident attributes to generate multiple event data groups. Dynamic time warping is performed on these event data groups, and hierarchical clustering is conducted based on the warping results to generate multiple typical behavioral pattern clusters. Frequent sequence mining is performed on these typical behavioral pattern clusters to extract event subsequence sets. Time-series distribution statistics are performed on these event subsequence sets to obtain event occurrence time distribution data and dwell time distribution data. These event occurrence time distribution data and dwell time distribution data are matched with the multiple typical behavioral pattern clusters to construct multiple group behavior templates. These multiple group behavior templates are indexed and stored to construct the group behavior baseline.

[0026] First, a set of historical behavioral event sequences for a predetermined number of consecutive days of residents in high-rise residential buildings is collected. Resident attribute vectors are established based on age group, family size, housing type, and daily routine. Historical behavioral event sequences with identical or similar attribute vectors are grouped into the same event data group. Then, for any two historical behavioral event sequences within each event data group, a dynamic time warping algorithm is used to calculate the time warped distance between the two sequences. A cumulative distance matrix is ​​established using the order of occurrence of behavioral events as a matching constraint. The warped distance corresponding to the path with the smallest cumulative distance is selected as the similarity between the two behavioral sequences. A distance matrix is ​​constructed based on the pairwise warped distances of all behavioral sequences. Further, a bottom-up hierarchical clustering algorithm is used, with the distance matrix as the clustering basis. The two behavioral clusters with the smallest distance are merged sequentially. Clustering stops when the distance between clusters exceeds a preset clustering threshold, generating multiple typical behavioral pattern clusters. Then, each typical behavioral pattern cluster is traversed, and a Prefix... The Span frequent sequence mining algorithm filters behavioral event subsequences that have consecutively appeared at a preset minimum support threshold, generating a set of event subsequences. It further analyzes the frequency of each event subsequence across different time periods within a day, calculates the probability of occurrence according to time intervals, and generates event occurrence time distribution data. Simultaneously, it analyzes the duration of each event subsequence, calculates the mean and standard deviation of each event duration, and generates dwell time distribution data. Then, it associates the event occurrence time distribution data and dwell time distribution data with corresponding typical behavioral pattern clusters to construct a group behavior template that includes behavioral event type, event probability, occurrence time distribution, dwell time distribution, and event transition relationships. Finally, it establishes a template index directory according to resident attribute categories, indexing and storing each group behavior template according to resident attribute category, behavioral pattern number, and template number, constructing a group behavior baseline for subsequent target object behavior template matching and invocation.

[0027] Furthermore, incremental learning updates are performed based on the initial individual behavior baseline of the target object to generate an individual behavior baseline. Methods include: Incremental learning updates are performed based on the initial individual behavior baseline of the target object to generate new behavioral events. When a new behavioral event arrives, the instantaneous anomaly score of the new behavioral event under the individual behavior baseline is calculated. A preset update threshold is set. If the instantaneous anomaly score is lower than the update threshold, the new behavioral event is subjected to exponential decay sliding updates to generate time distribution parameters and duration distribution parameters. Based on the time distribution parameters and duration distribution parameters, the new behavioral event is updated in multiple dimensions to generate the individual behavior baseline. If the instantaneous anomaly score is not lower than the update threshold, the update learning rate is reduced to generate an update learning rate to perform perturbation suppression on the new behavioral event, thereby obtaining the individual behavior baseline.

[0028] First, the structured behavioral event sequence generated in real time for the target object is sequentially integrated into the individual behavioral baseline according to the event occurrence time. Newly added structured behavioral events are defined as new behavioral events. When a new behavioral event arrives, the event occurrence time deviation, event duration deviation, and event transition deviation are calculated. The event occurrence time deviation is calculated using the standardized difference between the occurrence time of the new behavioral event and the mean occurrence time of the corresponding event in the individual behavioral baseline. The event duration deviation is calculated using the standardized difference between the duration of the new behavioral event and the mean duration of the corresponding event in the individual behavioral baseline. The event transition deviation is calculated using the absolute difference between the current event transition probability and the corresponding event transition probability in the individual behavioral baseline. Furthermore, all normal behavioral event samples of the target object within a preset number of consecutive days are statistically analyzed, and the variances corresponding to the event occurrence time deviation, event duration deviation, and event transition deviation are calculated. The weight of each deviation is determined based on its proportion to the total variance. in, Let these represent the sample variances corresponding to the event occurrence time deviation, event duration deviation, and event transition deviation, respectively, satisfying the following conditions: Then, the immediate anomaly score for the new behavioral event is calculated: in, These represent the deviations in event occurrence time, event duration, and event transfer, respectively; further, the instantaneous abnormality scores corresponding to all normal behavioral events within a consecutive preset number of days are statistically analyzed, and the mean of the abnormality scores is calculated. and standard deviation according to Determine the update threshold; when the instantaneous anomaly score meets the requirement... When a new behavioral event is determined to be a normal behavior, the event occurrence time distribution parameters and event duration distribution parameters are updated using an exponentially decaying sliding method. The event occurrence time distribution parameters are updated according to... Update the event duration distribution parameters according to... To be updated, among which, and These represent the mean of the time distribution and the mean of the duration before the update, respectively. Indicates the time when the new behavioral event occurred. Indicates the duration of the new behavioral event. This represents the basic update learning rate; after updating the time distribution parameters and duration distribution parameters, the event transition frequency is further recalculated based on the connection relationship between the new behavioral event and the preceding and following behavioral events, and the event transition probability matrix and event occurrence probability distribution are updated simultaneously to achieve multi-dimensional updates of the individual behavioral baseline; when the immediate anomaly score satisfies When a new behavioral event is determined to be potentially abnormal, the update learning rate is adjusted based on the ratio between the immediate abnormality score and the update threshold. in, The adjusted update learning rate is used to update the time distribution parameters, duration distribution parameters, event occurrence probability distribution, and event transition probability matrix. This ensures that abnormal behavior participates in the individual behavior baseline correction with only a low weight, thereby suppressing the perturbation of the individual behavior model by abnormal behavior and finally obtaining the updated individual behavior baseline.

[0029] The structured behavioral event sequence is compared with the individual behavioral baseline using a multi-dimensional deviation calculation to obtain a multi-dimensional anomaly deviation, which includes anomaly scores.

[0030] Furthermore, the multi-dimensional deviation of the structured behavioral event sequence from the individual behavioral baseline is calculated to obtain the multi-dimensional anomaly deviation. The method includes: Based on the structured behavioral event sequence, time dimension features, duration dimension features, spatial sequence dimension features, and event transition dimension features are extracted. The time dimension features are compared with the individual behavioral baseline to calculate the time distribution, obtaining a time deviation. The duration dimension features are compared with the duration distribution of the individual behavioral baseline to calculate the duration deviation. The spatial sequence dimension features are compared with the individual behavioral baseline to calculate the spatial trajectory sequence deviation. The event transition dimension features are compared with the individual behavioral baseline to calculate the event transition deviation. The time deviation, duration deviation, spatial trajectory sequence deviation, and event transition deviation are weighted and fused to obtain the multidimensional anomaly deviation.

[0031] First, the occurrence time, duration, spatial region sequence, and adjacent event transition relationships of the current behavioral event are extracted from the structured behavioral event sequence, serving as time-dimensional features, duration-dimensional features, spatial sequence-dimensional features, and event transition-dimensional features, respectively. Then, the baseline occurrence time mean, baseline occurrence time standard deviation, baseline duration mean, baseline duration standard deviation, baseline spatial trajectory sequence, and baseline event transition probability matrix for the corresponding event type are retrieved from the individual behavioral baseline. The time deviation is calculated based on the distribution of the current behavioral event occurrence time and the baseline occurrence time. Where t represents the time when the current behavioral event occurs. Indicates the average of the baseline occurrence time. Indicates the standard deviation of the baseline occurrence time. This represents a minimal constant to prevent the denominator from being zero; the duration deviation is calculated based on the distribution of the current event duration and the baseline duration. Where d represents the duration of the current behavioral event, This represents the average duration of the baseline. This represents the standard deviation of the baseline duration; the spatial trajectory sequence deviation is calculated based on the current spatial region sequence and the baseline spatial trajectory sequence, using edit distance for normalization. Where R represents the current spatial region sequence, This represents the baseline spatial trajectory sequence within the individual's behavioral baseline. Indicates edit distance; calculates event transition bias based on the event transition probability matrix in the individual behavior baseline and the current event transition relationship: in, Indicates event type Transfer to event type The baseline transition probability is then calculated; subsequently, the fusion weights are calculated based on the historical statistical stability of each dimension in the individual behavioral baseline, where the standard deviations of time deviation, duration deviation, spatial trajectory sequence deviation, and event transition deviation in the historical normal behavior sample are respectively denoted as . Calculate the initial weights for each dimension separately: The initial weights for each dimension were then normalized. Finally, the multidimensional anomaly deviation is calculated based on the normalized fusion weights: in, It represents the multidimensional anomaly deviation and is used as an anomaly score output for subsequent multi-level early warning determination.

[0032] Based on the aforementioned abnormal score, multi-level early warning conditions are triggered, and multi-level abnormal behavior early warning information is generated to provide early warning of abnormal behavior of people in high-rise residential buildings.

[0033] Furthermore, such as Figure 2 As shown, based on the aforementioned abnormal scoring triggering multi-level early warning conditions, multi-level abnormal behavior early warning information is generated to provide early warning of abnormal behavior of people in high-rise residential buildings. The method includes: A first-level scoring threshold and a first-level duration are set. When the abnormal score continuously exceeds the first-level scoring threshold within the first-level duration, a first warning message is generated to alert residents of high-rise residential buildings regarding abnormal behavior. A second-level scoring threshold and a specific abnormal event type combination are set. When the abnormal score exceeds the second-level scoring threshold and the specific abnormal event type combination is simultaneously identified, a second warning message is generated to alert residents of high-rise residential buildings regarding abnormal behavior. Vulnerability tags of target objects are obtained, and the warning sensitivity is dynamically adjusted based on the vulnerability tags. A third-level scoring threshold is set. Based on the warning sensitivity, it is determined whether the vulnerability tag indicates a high-risk object. When the vulnerability tag indicates a high-risk object and the abnormal score exceeds the third-level scoring threshold, an emergency warning message is triggered and simultaneously pushed to a preset management terminal to alert residents of high-rise residential buildings regarding abnormal behavior.

[0034] Preferably, a first-level scoring threshold is pre-set. Level 1 duration Secondary scoring threshold and the three-level scoring threshold An early warning rule base is established; the system continuously acquires anomaly scores of target objects at a preset sampling period, and when the continuous sampling time reaches the first-level duration... At that time, all abnormal scores within the statistical duration window meet the requirement of being greater than the first-level score threshold. If the conditions for a Level 1 warning are met, a warning message is generated. The first warning message must include at least the target object's identity, anomaly score, anomaly event type, time of occurrence, floor, and current location, and is sent to the property management terminal. If the anomaly score continues to rise and exceeds the Level 2 warning threshold, The system then invokes a preset abnormal event combination rule base to perform continuous event matching on the current structured behavioral event sequence. When a preset abnormal event type combination is identified, it is determined that the conditions for a level two warning are met. This abnormal event type combination refers to any one or more combinations of the following occurring within a preset time period: loitering, abnormal elevator stoppage, prolonged nighttime stay in restricted areas, and frequent cross-floor travel. The system generates a verification warning message. The second warning message includes the content of the first warning message, as well as the abnormal behavior trajectory, associated event sequence, and corresponding monitoring screen index, and is simultaneously pushed to the property management terminal and the security personnel's mobile terminal. Furthermore, the system reads the vulnerability tags of the target object and configures different warning sensitivity coefficients based on different vulnerability tags. The elderly, children, people with mobility impairments, and people living alone for extended periods correspond to higher sensitivity coefficients. The system then dynamically adjusts the level three scoring threshold based on these sensitivity coefficients. in, This indicates the revised three-level rating threshold. This represents the early warning sensitivity coefficient, and If the target object is a high-risk object and the abnormal score exceeds the revised level three scoring threshold. If the conditions for a Level 3 warning are met, an emergency warning message is generated. Simultaneously, the emergency warning message, the real-time location of the target object, its behavior trajectory within the most recent preset time period, related video clips, and anomaly scores are pushed to the preset management terminal. At the same time, a continuous tracking mechanism is activated to update the real-time location of the target object and the anomaly scores at preset time intervals until the management personnel confirm that the handling has been completed and the warning process ends.

[0035] In summary, the embodiments of this application have at least the following technical effects: First, the target objects in high-rise residential buildings are located and subjected to real-time multi-source perception fusion to generate a fused behavioral data stream. Next, the fused behavioral data stream is analyzed for spatiotemporal context to extract structured behavioral event sequences of the target objects. Further, based on the structured behavioral event sequences, group behavior analysis is performed to establish an initial individual behavioral baseline for the target objects. Incremental learning and updates are then performed according to this initial baseline to generate individual behavioral baselines. Then, multi-dimensional deviation calculations are performed between the structured behavioral event sequences and the individual behavioral baselines to obtain a multi-dimensional anomaly deviation, which includes an anomaly score. Finally, multi-level early warning conditions are triggered based on the anomaly score to generate multi-level abnormal behavior early warning information for alerting residents in high-rise residential buildings to report abnormal behaviors. This approach solves the technical problems of low accuracy and high false alarm / missed alarm rates in existing technologies for identifying abnormal behaviors in high-rise residential buildings, achieving accurate identification of abnormal behaviors and reducing the false alarm rate.

[0036] Example 2 is based on the same inventive concept as the method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines in the previous examples, such as... Figure 3As shown, this application provides an early warning system for abnormal behavior of people in high-rise residential buildings based on behavioral baselines, wherein the system includes: Data fusion module 11: Locates target objects in high-rise residential buildings and performs real-time multi-source perception fusion to generate a fused behavior data stream; Behavior analysis module 12: Performs spatiotemporal context behavior analysis on the fused behavior data stream to extract the structured behavior event sequence of the target objects; Behavior analysis module 13: Performs group behavior analysis based on the structured behavior event sequence, sets an initial individual behavior baseline for the target objects, and performs incremental learning and updating according to the initial individual behavior baseline to generate an individual behavior baseline; Deviation calculation module 14: Performs multi-dimensional deviation calculation between the structured behavior event sequence and the individual behavior baseline to obtain a multi-dimensional abnormal deviation degree, which includes an abnormal score; Early warning module 15: Triggers multi-level early warning conditions based on the abnormal score to generate multi-level abnormal behavior early warning information to warn of abnormal behavior of people in high-rise residential buildings.

[0037] Furthermore, the data fusion module 11 is used to perform the following methods: Multi-view synchronous acquisition of public areas in high-rise residential buildings generates multi-view video streams, simultaneously acquiring access control event streams, floor infrared trigger data, and point cloud data of the public areas. Multi-target re-identification is performed on the multi-view video streams to obtain cross-camera trajectory segments. Target clustering and tracking are performed on the point cloud data to obtain radar target trajectories. The access control event streams and cross-camera trajectory segments are timestamped to generate target object identification identifiers. Continuous trajectory identification is performed by traversing the radar target trajectories to determine missing trajectory data segments. Data interpolation is performed based on the missing trajectory data segments to complete the continuous full trajectory. Spatial constraint registration is performed between the continuous full trajectory and the floor infrared trigger data to generate a multimodal time series, which is then used as the fused behavioral data stream.

[0038] Furthermore, the behavior parsing module 12 is used to perform the following methods: A spatial topology of a high-rise residential building is constructed. Based on this spatial topology and the fused behavioral data stream, a spatiotemporal location map of the target object is constructed. A temporal graph attention network is used to extract features from the spatiotemporal location map, generating a sequence of behavioral node embedding vectors. Behavior recognition is performed based on the sequence of behavioral node embedding vectors to identify and determine the atomic behavior categories of the target object, obtaining a sequence of atomic behavior categories. A temporal segmentation network is used to detect event boundaries in the sequence of atomic behavior categories, determining multiple behavioral event boundary data, which include event start time and event end time. Feature extraction is performed according to the event start time, the event end time, and the sequence of atomic behavior categories to obtain spatiotemporal attribute features and object association features. The structured behavioral event sequence is constructed based on the spatiotemporal attribute features and the object association features.

[0039] Furthermore, the behavior parsing module 12 is used to perform the following methods: Based on the spatial topology of the high-rise residential building, functional zones are defined, and multiple functional areas are delineated. The fusion behavior data stream is traversed for location matching and identification to determine multiple location points. Timestamps are extracted from the fusion behavior data stream, and the multiple location points are mapped to the multiple functional areas according to the timestamps to construct region-time tuple nodes. Based on the region-time tuple nodes, the functional area change analysis of the target object between adjacent time slices is performed, and multiple directed edges of nodes are established. The multiple directed edges of nodes are added to the spatiotemporal location map.

[0040] Furthermore, the behavior analysis module 13 is used to perform the following methods: A group behavior baseline is constructed, and the attribute features of the target object are extracted and matched with the group behavior baseline to determine a group behavior template, which includes a set of frequent sequence patterns. Attention distribution encoding is performed based on the set of frequent sequence patterns to construct a group behavior prototype vector. An individual behavior space is constructed by mapping the group behavior prototype vector to the individual behavior space using the structured behavior event sequence as a constraint, and calculating the initial event probability distribution data, the initial event duration distribution data, and the initial event transition probability matrix. The initial event probability distribution data, the initial event duration distribution data, and the initial event transition probability matrix are integrated to construct the initial individual behavior baseline.

[0041] Furthermore, the behavior analysis module 13 is used to perform the following methods: A collection of historical behavioral event sequences of residents in high-rise residential buildings is gathered. These sequences are then grouped according to resident attributes to generate multiple event data groups. Dynamic time warping is performed on these event data groups, and hierarchical clustering is conducted based on the warping results to generate multiple typical behavioral pattern clusters. Frequent sequence mining is performed on these typical behavioral pattern clusters to extract event subsequence sets. Time-series distribution statistics are performed on these event subsequence sets to obtain event occurrence time distribution data and dwell time distribution data. These event occurrence time distribution data and dwell time distribution data are matched with the multiple typical behavioral pattern clusters to construct multiple group behavior templates. These multiple group behavior templates are indexed and stored to construct the group behavior baseline.

[0042] Furthermore, the behavior analysis module 13 is used to perform the following methods: Incremental learning updates are performed based on the initial individual behavior baseline of the target object to generate new behavioral events. When a new behavioral event arrives, the instantaneous anomaly score of the new behavioral event under the individual behavior baseline is calculated. A preset update threshold is set. If the instantaneous anomaly score is lower than the update threshold, the new behavioral event is subjected to exponential decay sliding updates to generate time distribution parameters and duration distribution parameters. Based on the time distribution parameters and duration distribution parameters, the new behavioral event is updated in multiple dimensions to generate the individual behavior baseline. If the instantaneous anomaly score is not lower than the update threshold, the update learning rate is reduced to generate an update learning rate to perform perturbation suppression on the new behavioral event, thereby obtaining the individual behavior baseline.

[0043] Furthermore, the deviation calculation module 14 is used to perform the following method: Based on the structured behavioral event sequence, time dimension features, duration dimension features, spatial sequence dimension features, and event transition dimension features are extracted. The time dimension features are compared with the individual behavioral baseline to calculate the time distribution, obtaining a time deviation. The duration dimension features are compared with the duration distribution of the individual behavioral baseline to calculate the duration deviation. The spatial sequence dimension features are compared with the individual behavioral baseline to calculate the spatial trajectory sequence deviation. The event transition dimension features are compared with the individual behavioral baseline to calculate the event transition deviation. The time deviation, duration deviation, spatial trajectory sequence deviation, and event transition deviation are weighted and fused to obtain the multidimensional anomaly deviation.

[0044] Furthermore, the first processing module 12 is used to perform the following methods: A first-level scoring threshold and a first-level duration are set. When the abnormal score continuously exceeds the first-level scoring threshold within the first-level duration, a first warning message is generated to alert residents of high-rise residential buildings regarding abnormal behavior. A second-level scoring threshold and a specific abnormal event type combination are set. When the abnormal score exceeds the second-level scoring threshold and the specific abnormal event type combination is simultaneously identified, a second warning message is generated to alert residents of high-rise residential buildings regarding abnormal behavior. Vulnerability tags of target objects are obtained, and the warning sensitivity is dynamically adjusted based on the vulnerability tags. A third-level scoring threshold is set. Based on the warning sensitivity, it is determined whether the vulnerability tag indicates a high-risk object. When the vulnerability tag indicates a high-risk object and the abnormal score exceeds the third-level scoring threshold, an emergency warning message is triggered and simultaneously pushed to a preset management terminal to alert residents of high-rise residential buildings regarding abnormal behavior.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines, characterized in that, The method includes: The system locates target objects in high-rise residential buildings and performs real-time multi-source perception fusion to generate a fused behavioral data stream. Spatiotemporal context behavior parsing is performed on the fused behavior data stream to extract the structured behavior event sequence of the target object; Group behavior analysis is performed based on the structured behavioral event sequence. An initial individual behavior baseline of the target object is set, and incremental learning and updating are performed according to the initial individual behavior baseline of the target object to generate an individual behavior baseline. The structured behavioral event sequence is compared with the individual behavioral baseline in a multidimensional deviation calculation to obtain a multidimensional anomaly deviation, which includes anomaly scores. Based on the aforementioned abnormal score, multi-level early warning conditions are triggered, and multi-level abnormal behavior early warning information is generated to provide early warning of abnormal behavior of people in high-rise residential buildings.

2. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 1, characterized in that, The method for locating target objects in high-rise residential buildings and performing real-time multi-source sensing fusion to generate a fused behavioral data stream includes: Multi-view synchronous data acquisition is performed on the public areas of high-rise residential buildings to generate multi-view video streams. Access control event streams, floor infrared trigger data, and point cloud data of the public areas of high-rise residential buildings are also collected synchronously. Perform multi-target re-identification on the multi-view video stream to obtain cross-camera trajectory segments; Target clustering and tracking are performed on the point cloud data to obtain radar target trajectories; The access control event stream is timestamped and the cross-camera trajectory segment to generate a target object identity identifier. The radar target trajectory is traversed to perform continuous trajectory identification, the missing trajectory data segment is determined, and data interpolation is performed to complete the missing trajectory data segment to obtain the continuous full trajectory. The continuous full trajectory is spatially constrained and registered with the floor infrared trigger data to generate a multimodal time series, which is then used as the fused behavioral data stream.

3. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 1, characterized in that, The spatiotemporal context behavior parsing of the fused behavior data stream is performed to extract the structured behavior event sequence of the target object. The method includes: Construct a spatial topology of a high-rise residential building, and based on the spatial topology and the fused behavioral data stream, construct a spatiotemporal location map of the target object; A temporal graph attention network is used to extract features from the spatiotemporal location graph to generate a sequence of behavior node embedding vectors; Based on the behavior node embedding vector sequence, behavior recognition is performed to identify and determine the atomic behavior categories of the target object, thereby obtaining an atomic behavior category sequence. A temporal segmentation network is used to perform event boundary detection on the atomic behavior category sequence to determine multiple behavior event boundary data, which include event start time and event end time. Feature extraction is performed based on the event start time, the event end time, and the atomic behavior category sequence to obtain spatiotemporal attribute features and object association features; The structured behavioral event sequence is constructed based on the spatiotemporal attribute features and the object association features.

4. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 3, characterized in that, Based on the spatial topology and the fused behavioral data stream, a spatiotemporal location map of the target object is constructed, the method including: Based on the spatial topology of the high-rise residential building, functional zones are defined, and multiple functional areas are delineated. The fused behavioral data stream is traversed to perform location matching and identification, and multiple location points are determined. Based on the timestamp extracted from the fused behavior data stream, the multiple location points are mapped to the multiple functional areas according to the timestamp, and a region-time tuple node is constructed. Based on the aforementioned region-time tuple nodes, the functional region change analysis of the target object between adjacent time slices is performed, and multiple directed edges of nodes are established. Add the directed edges of the multiple nodes to the spatiotemporal location graph.

5. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 1, characterized in that, Based on the structured behavioral event sequence, group behavior analysis is performed, and an initial individual behavioral baseline for the target group is established. The method includes: Construct a group behavior baseline, extract the attribute features of the target object and match them with the group behavior baseline to determine a group behavior template, wherein the group behavior template contains a set of frequent sequence patterns; Attention distribution encoding is performed based on the frequent sequence pattern set to construct a group behavior prototype vector; Construct an individual behavior space, use the structured behavior event sequence as a constraint to map the group behavior prototype vector to the individual behavior space, and calculate the initial event probability distribution data, the initial event duration distribution data, and the initial event transition probability matrix; The initial event probability distribution data, the initial event duration distribution data, and the initial event transition probability matrix are integrated to construct the initial individual behavior baseline.

6. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 5, characterized in that, The process and methods for constructing a group behavior baseline include: The historical behavioral event sequence set of residents in high-rise residential buildings is collected, and the historical behavioral event sequence set is grouped according to resident attributes to generate multiple event data groups; Dynamic time warping is performed based on the multiple event data groups, and hierarchical clustering is performed based on the warping results to generate multiple typical behavior pattern clusters. The multiple typical behavior pattern clusters are traversed to perform frequent sequence mining and extract event subsequence sets. Based on the event subsequence set, time-series distribution statistics are performed to obtain event occurrence time distribution data and dwell time distribution data; Multiple group behavior templates are constructed by matching the event occurrence time distribution data, the dwell time distribution data, and the multiple typical behavior pattern clusters. The multiple group behavior templates are indexed and stored to construct the group behavior baseline.

7. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 1, characterized in that, Incremental learning and updating are performed based on the initial individual behavior baseline of the target object to generate the individual behavior baseline. Methods include: Incremental learning and updates are performed based on the initial individual behavioral baseline of the target object to generate new behavioral events; When the new behavioral event arrives, calculate the immediate abnormality score of the new behavioral event under the individual's behavioral baseline; A preset update threshold is set. If the instantaneous anomaly score is lower than the update threshold, the new behavioral event is subjected to exponential decay sliding update to generate time distribution parameters and duration distribution parameters. Based on the time distribution parameters and the duration distribution parameters, the new behavioral events are updated in multiple dimensions to generate individual behavioral baselines. If the instantaneous anomaly score is not lower than the update threshold, the update learning rate is reduced, and an update learning rate is generated to perturb and suppress the new behavioral event, thereby obtaining the individual behavioral baseline.

8. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 1, characterized in that, The multidimensional deviation of the structured behavioral event sequence from the individual behavioral baseline is calculated to obtain the multidimensional anomaly deviation. The method includes: Based on the structured behavioral event sequence, extract time dimension features, duration dimension features, spatial sequence dimension features, and event transition dimension features; The time distribution is calculated based on the time dimension characteristics and the individual behavior baseline to obtain the time deviation degree; The duration deviation is calculated based on the duration dimension characteristics and the duration distribution of the individual behavioral baseline. The spatial trajectory sequence deviation is calculated based on the spatial sequence dimensional features and the individual behavioral baseline. The event transfer deviation is calculated based on the event transfer dimension features and the individual behavior baseline. The time deviation, duration deviation, spatial trajectory sequence deviation, and event transfer deviation are weighted and fused to obtain the multidimensional anomaly deviation.

9. The method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in claim 1, characterized in that, Based on the aforementioned abnormal scoring triggering multi-level early warning conditions, multi-level abnormal behavior early warning information is generated to provide early warning of abnormal behavior of residents in high-rise residential buildings. The method includes: A first-level scoring threshold and a first-level duration are set. When the abnormal score continuously exceeds the first-level scoring threshold within the first-level duration, a warning message is triggered to generate the first warning message to warn of abnormal behavior of people in high-rise residential buildings. A secondary scoring threshold and a specific combination of abnormal event types are set. When the abnormal score exceeds the secondary scoring threshold and the specific combination of abnormal event types is simultaneously identified, a verification warning message is triggered to generate a second warning message to warn of abnormal behavior of people in high-rise residential buildings. Obtain the vulnerability tags of the target object, and dynamically adjust the warning sensitivity based on the vulnerability tags; A three-level scoring threshold is set. Based on the warning sensitivity, it is determined whether the vulnerability label is a high-risk object. When the vulnerability label is a high-risk object and the abnormal score exceeds the three-level scoring threshold, an emergency warning message is triggered and the emergency warning message is pushed to the preset management terminal to warn of abnormal behavior of people in high-rise residential buildings.

10. A high-rise residential building occupant abnormal behavior early warning system based on behavioral baselines, characterized in that: For implementing the method for early warning of abnormal behavior of people in high-rise residential buildings based on behavioral baselines as described in any one of claims 1-9, the system comprises: Data fusion module: Locates target objects in high-rise residential buildings, performs real-time multi-source sensing fusion, and generates a fused behavior data stream; Behavior parsing module: performs spatiotemporal context behavior parsing on the fused behavior data stream to extract the structured behavior event sequence of the target object; Behavior analysis module: Based on the structured behavioral event sequence, perform group behavior analysis, set an initial individual behavior baseline for the target object, perform incremental learning and updating according to the initial individual behavior baseline of the target object, and generate individual behavior baselines; Deviation calculation module: Performs multi-dimensional deviation calculation between the structured behavioral event sequence and the individual behavioral baseline to obtain multi-dimensional abnormal deviation degree, which includes anomaly score; Early warning module: Based on the abnormal score, trigger multi-level early warning conditions to generate multi-level abnormal behavior early warning information to provide early warning of abnormal behavior of people in high-rise residential buildings.

Images