A monitoring video data abnormal behavior detection method based on behavior feature recognition

Abstract

The application discloses a kind of based on behavior characteristic identification's monitoring video data abnormal behavior detection method, comprising: continuous monitoring video data is collected, forms video frame sequence;Perform target detection and target consistency association processing, form behavior characteristic time sequence;Improved CMAC network is constructed, and the corresponding behavior response output sequence is generated;In normal and abnormal reference period, corresponding statistical parameters are extracted, combined to form hypothesis parameter generation sequence;Perform sequential probability ratio test, form final cumulative statistic;Perform abnormality determination, and output abnormal behavior detection result.The application realizes the continuous online detection and early determination of abnormal behavior in monitoring video by introducing improved CMAC network and sequential probability ratio test.

Description

Technical Field

[0001] This invention relates to the field of behavioral feature recognition technology, and in particular to a method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition. Background Technology

[0002] With the continuous development of intelligent monitoring systems, abnormal behavior detection technology based on video data has been widely applied in scenarios such as public place security, traffic monitoring, and industrial production supervision. Traditional technologies typically rely on surveillance cameras to collect continuous video and analyze the behavior of people or objects within the monitored area. Traditional solutions identify and alarm on abnormal behavior by setting behavioral thresholds, statistical rules, or classification models. However, due to the complexity of the monitoring environment, the diversity of behaviors, and video noise interference, practical applications often face high false alarm rates and insensitivity to short-term anomalies.

[0003] Existing technologies attempt to introduce statistical discrimination methods to perform time-series analysis of behavioral characteristics and use probabilistic statistical models to distinguish between normal and abnormal behaviors. Current methods typically rely on assumed distribution parameters to accumulate statistics on behavioral responses and make a judgment when the statistics reach a preset threshold, thus possessing a certain online detection capability. However, existing basic solutions mostly use fixed threshold accumulation methods, which still have limited ability to characterize the gradual changes in behavioral responses over time. When facing scenarios with frequent behavioral fluctuations or diverse abnormal patterns, problems such as unstable judgment, delayed response, or statistical accumulation failure still exist.

[0004] Therefore, how to provide a method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an abnormal behavior detection method for surveillance video data based on behavioral feature recognition. This invention comprehensively utilizes video target detection and tracking, behavioral feature temporal modeling, and sequential statistical discrimination methods to continuously and online determine the abnormal states of behavioral objects in surveillance videos. This invention constructs a target trajectory sequence of behavioral objects, extracts behavioral features from it to form a behavioral feature time series, introduces an improved CMAC network to quantize and map the behavioral features and model memory responses, generating a stable behavioral response output sequence. Combined with a sequential probability ratio test method, the behavioral responses are statistically accumulated hourly to complete the anomaly determination. This invention can continuously accumulate abnormal evidence during the gradual evolution of behavior, achieving early identification and stable determination of abnormal behavior. It has advantages such as low false alarm rate, timely response, and strong adaptability to complex behavioral changes, making it suitable for abnormal behavior detection applications in complex monitoring scenarios.

[0006] A method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to an embodiment of the present invention includes: Collect continuous monitoring video data and perform time sampling on the continuous monitoring video data according to a unified time base to form a video frame sequence; The video frame sequence is subjected to target detection and target consistency association processing to determine the behavioral object corresponding to each time sampling point and construct the corresponding target trajectory sequence. Behavioral features are extracted at each time sampling point to form a behavioral feature time series. Based on behavioral feature time series, an improved CMAC network is constructed to perform input space quantization mapping on the behavioral features at each time sampling point, activate the corresponding memory unit set, and generate the corresponding behavioral response output sequence. Based on the behavioral response output sequence, the corresponding statistical parameters of the behavioral response output sequence are extracted in the order of time sampling points in the normal and abnormal reference periods, and combined to form the hypothesis parameter generation sequence. Generate a sequence based on the hypothetical parameters, perform a sequential probability ratio test, construct a corresponding statistical observation for the behavioral response output value at each time sampling point, calculate the corresponding likelihood ratio and perform sequential cumulative processing to form the final cumulative statistic. Anomaly detection is performed based on the final cumulative statistics, and the abnormal behavior detection results are output.

[0007] Optionally, forming the video frame sequence includes: Acquire continuous monitoring video data from surveillance camera equipment, record the time information corresponding to the continuous monitoring video data, and determine the start time point of video acquisition; Based on a unified time sampling rule set at the start time point, multiple time sampling points are generated on the time axis of continuous monitoring video data at fixed time intervals. For each time sampling point, the video frame closest to the time of the time sampling point is selected from the continuous monitoring video data and used as the video frame corresponding to the current time sampling point. The video frames corresponding to each time sampling point are arranged in chronological order to form a video frame sequence arranged in chronological order.

[0008] Optionally, constructing the corresponding target trajectory sequence includes: Perform target detection processing on each video frame in the video frame sequence, obtain the target candidate set of all detected targets in the video frame, and record the target position parameters and target appearance feature parameters for each detected target in the video frame; Based on video frame sequences, target consistency association processing is performed on the target candidate sets in two adjacent video frames. The target matching value is calculated based on the target position parameters and target appearance feature parameters to form cross-frame target association results. Based on the cross-frame target association results, a unified behavior object identifier is assigned to the target candidates on the same association link, and the behavior object corresponding to the behavior object identifier is determined at each time sampling point of the video frame sequence, forming a time sampling point-level behavior object sequence. For each behavior object identifier, the target position parameters of the behavior object at each time sampling point are aggregated to construct the target trajectory sequence of the behavior object at continuous time sampling points.

[0009] Optionally, the formation of the behavioral feature time series includes: Based on the target trajectory sequence, for each action object, the corresponding target position parameters are read at the current time and the previous time sampling point, the displacement between adjacent time sampling points is calculated, and the velocity, acceleration and motion direction angle are calculated based on the displacement and a fixed time interval. The direction change is calculated based on the motion direction angle of adjacent time sampling points, and the trajectory curvature feature is calculated based on the direction change and displacement. At the same time, the presence of a stationary state is determined based on the velocity and velocity threshold. The number of time sampling points that continuously meet the stationary state is accumulated to obtain the stationary duration feature. The displacement, velocity, acceleration, motion direction angle, direction change, trajectory curvature features, and dwell time features obtained at the current time sampling point are combined in sequence to form the behavior feature vector corresponding to the current time sampling point, thus obtaining the behavior feature time series.

[0010] Optionally, generating the corresponding behavioral response output sequence includes: An improved CMAC network is constructed. For each time sampling point in the behavioral feature time series, the behavioral features of the current time sampling point are divided according to the granularity level of the behavioral features to obtain a multi-granularity feature set. For each granularity level in the multi-granularity feature set, the input space partitioning rules and quantization unit set corresponding to the granularity level are configured respectively, and the quantization interval numbering rules and quantization address encoding rules are configured for each granularity level respectively, forming an input space quantization configuration that corresponds one-to-one with each granularity level; For each time sampling point, based on the input space quantization configuration, the input space quantization mapping is performed one by one for each granularity level feature to obtain the quantization interval number set corresponding to each component of the granularity level feature, and the quantization address code is generated based on the quantization interval number set, while the center index position corresponding to the quantization address code is determined. Memory indexes are generated based on the quantization address encoding corresponding to each granularity level, and the activated memory cell subsets are determined in the memory cell set of each granularity level. At the same time, the relative positional relationship of the activated memory cell subsets in the quantization space is converted into structural offset information, and the structural offset information is recorded as relative offset labels to construct a position-sensitive structural memory matrix. Based on the position-sensitive structural memory matrix, the weight storage structure is set as a tensor parameter structure. Tensorized weight indexing is performed on the tensor parameter structure based on the memory index to obtain the intermediate response output corresponding to each granularity level. Position-aware attenuation weights are generated based on the relative distance between the features of each granular level and the corresponding center index position. The intermediate response outputs corresponding to each granular level are weighted and synthesized to obtain the behavioral response output values ​​corresponding to the time sampling points, and a behavioral response output sequence is generated.

[0011] Optionally, the combination forms a hypothesis parameter generation sequence, including: Based on the temporal order of the time sampling points, establish corresponding time sampling point identifiers for each behavioral response output value in the behavioral response output sequence, forming a behavioral response output index sequence aligned with time sampling points; In the behavior response output index sequence, determine the normal reference period and the abnormal reference period, and denote the set of behavior response output values ​​within the normal reference period as the normal reference set, and denote the set of behavior response output values ​​within the abnormal reference period as the abnormal reference set. Based on the normal reference set, the corresponding statistical parameter entries are calculated according to the time sampling point order, and the statistical parameter entries are arranged according to the time sampling point order to form a normal behavior hypothesis parameter sequence. Each statistical parameter entry includes the mean parameter and variance parameter of the corresponding time sampling point. Based on the abnormal reference set, the corresponding statistical parameter entries are calculated according to the time sampling point order, and the statistical parameter entries are arranged according to the time sampling point order to form the abnormal behavior hypothesis parameter sequence. Each statistical parameter entry includes the mean parameter and variance parameter of the corresponding time sampling point. The normal behavior hypothesis parameter sequence and the abnormal behavior hypothesis parameter sequence are combined in the order of time sampling points to obtain the hypothesis parameter generation sequence.

[0012] Optionally, the formation of the final cumulative statistic includes: Generate a sequence based on the hypothesis parameters, and read the corresponding behavioral response output value and the corresponding normal behavior hypothesis parameters and abnormal behavior hypothesis parameters for each time sampling point; For each time sampling point, based on the normal behavior hypothesis parameters and the abnormal behavior hypothesis parameters, the probability density values ​​of the current behavior response output value under the normal hypothesis period and the abnormal hypothesis period are calculated respectively. The ratio of the abnormal hypothesis probability density value to the normal hypothesis probability density value is used as the likelihood ratio of the time sampling point. The logarithm of the likelihood ratio is taken to obtain the log-likelihood increment. Sequential accumulation is performed based on the log-likelihood increment, and a directional gain coefficient is set according to the sign relationship between the log-likelihood increment of the current time sampling point and the previous time sampling point. After weighting the log-likelihood increment of the current time sampling point, the initial cumulative statistic is obtained. A threshold adjustment factor is generated based on the variation amplitude between the behavioral response output values ​​of adjacent time sampling points and the normal hypothesis probability density value. The basic upper limit threshold and the basic lower limit threshold are updated according to the threshold adjustment factor to obtain the dynamic upper limit threshold and the dynamic lower limit threshold corresponding to the current time sampling point. Construct an abnormal pattern library. For the current time sampling point, generate the pattern matching probability based on the matching results between the behavior response output subsequence containing the current time sampling point and each abnormal pattern template. The update magnitude of the initial cumulative statistics corresponding to the current time sampling point is adjusted based on the pattern matching probability. Sequential accumulation is performed on the log-likelihood increment based on two different time scales to form a multi-time-scale cumulative statistical trajectory. The initial cumulative statistics are weighted and fused based on the consistency results between the cumulative statistical trajectories of each time scale to obtain the final cumulative statistics. During the sequential accumulation process, the behavioral response entropy value is calculated based on the probability density values ​​of the normal hypothesis and the probability density values ​​of the abnormal hypothesis, and the behavioral response entropy value is accumulated. When the accumulated behavioral response entropy value reaches the entropy threshold, the final accumulated statistic is reset; otherwise, the final accumulated statistic, the dynamic upper bound threshold, the dynamic lower bound threshold, and the early judgment flag are output.

[0013] Optionally, the output of abnormal behavior detection results includes: At each sampling point, the final cumulative statistic is compared with the dynamic upper bound threshold and the dynamic lower bound threshold. When the advance judgment flag is valid, an anomaly judgment candidate result is directly generated. When the advance judgment flag is invalid, an anomaly judgment candidate result is generated based on the comparison result. For the pending states in the anomaly detection candidate results, a consistency judgment value is generated based on the consistency results of pattern matching probability and multi-timescale cumulative statistical trajectory. Confirmation and cancellation operations are performed on the anomaly detection candidate results to obtain the final anomaly detection result corresponding to the time sampling point. When the final anomaly determination result is an anomaly, the abnormal behavior detection result corresponding to the time sampling point is output.

[0014] The beneficial effects of this invention are: This invention enables continuous and stable detection of abnormal behavior. First, it constructs a video frame sequence arranged chronologically. Then, through target detection and cross-frame consistency correlation processing, it obtains a target trajectory sequence of the behavioral object at continuous time sampling points. Behavioral features such as displacement, velocity, acceleration, direction change, trajectory curvature, and dwell time are extracted from the target trajectory to form a behavioral feature time series.

[0015] This invention introduces an improved CMAC network to perform multi-granularity quantitative mapping and memory response modeling on behavioral feature time series, generating behavioral response output sequences corresponding to continuous time sampling points. This method can transform complex, continuously changing behavioral features into structured behavioral response representations, enhancing the ability to characterize different behavioral scales and change patterns. A sequential probability ratio test method is used to accumulate the behavioral response at each time sampling point to obtain the corresponding cumulative statistics.

[0016] This invention continuously accumulates abnormal evidence during the gradual evolution of behavior, avoiding false alarms caused by one-time judgments or fixed threshold methods, and achieving early identification and accurate confirmation of abnormal behavior. In complex monitoring environments, this invention offers advantages such as a continuous judgment process, strong adaptability to behavioral changes, and timely response to the initial occurrence of anomalies. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition, as proposed in this invention. Figure 2 This is a flowchart illustrating the execution of an improved CMAC network for an abnormal behavior detection method in surveillance video data based on behavioral feature recognition, as proposed in this invention. Figure 3 This is a flowchart illustrating the execution of the sequential probability ratio test in a method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition, as proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition, comprising: Collect continuous monitoring video data and perform time sampling on the continuous monitoring video data according to a unified time base to form a video frame sequence; The video frame sequence is subjected to target detection and target consistency association processing to determine the behavioral object corresponding to each time sampling point and construct the corresponding target trajectory sequence. Behavioral features are extracted at each time sampling point to form a behavioral feature time series. Based on behavioral feature time series, an improved CMAC network is constructed to perform input space quantization mapping on the behavioral features at each time sampling point, activate the corresponding memory unit set, and generate the corresponding behavioral response output sequence. Based on the behavioral response output sequence, the corresponding statistical parameters of the behavioral response output sequence are extracted in the order of time sampling points in the normal and abnormal reference periods, and combined to form the hypothesis parameter generation sequence. Generate a sequence based on the hypothetical parameters, perform a sequential probability ratio test, construct a corresponding statistical observation for the behavioral response output value at each time sampling point, calculate the corresponding likelihood ratio and perform sequential cumulative processing to form the final cumulative statistic. Anomaly detection is performed based on the final cumulative statistics, and the abnormal behavior detection results are output.

[0020] In this embodiment, forming a video frame sequence includes: Acquire continuous monitoring video data from surveillance camera equipment, record the time information corresponding to the continuous monitoring video data, and determine the start time point of video acquisition; Based on a unified time sampling rule set from the starting time point, multiple time sampling points are generated on the timeline of continuously monitored video data at fixed time intervals, where: The unified time sampling rule based on the start time point is specifically as follows: The moment when the surveillance camera starts collecting video data is taken as the starting reference point of the time axis. The time axis is divided into time positions in a sequential manner according to equal time intervals, and each division position corresponds to a unique time sampling point. The fixed time interval is set to one second. Starting from the initial time point, a new time sampling point is generated every second. For each time sampling point, the video frame closest to the time of the time sampling point is selected from the continuous monitoring video data and used as the video frame corresponding to the current time sampling point. The video frames corresponding to each time sampling point are arranged in chronological order to form a video frame sequence arranged in chronological order.

[0021] In this embodiment, constructing the corresponding target trajectory sequence includes: Target detection processing is performed on each video frame in the video frame sequence to obtain a target candidate set of all detected targets in the video frame, and the target position parameters and target appearance feature parameters of each detected target are recorded in the video frame. Specifically, the target detection processing on each video frame in the video frame sequence involves: The video frame image data is read frame by frame according to the time order of the video frame in the video frame sequence. During the traversal, the video frame is continuously divided into image regions of fixed size. When the image region simultaneously meets the conditions that the brightness change exceeds the brightness discrimination threshold, the edge change density exceeds the edge discrimination threshold, and the pixel difference with the adjacent image region exceeds the region difference discrimination threshold, the current image region is determined to be an image region containing the target, and the corresponding spatial location range is recorded as the target location parameter. The appearance feature parameters are extracted based on the brightness distribution, color distribution and edge structure distribution of the pixels in the image region. Based on the video frame sequence, target consistency association processing is performed on the target candidate set in two adjacent video frames. Target matching values ​​are calculated based on target location parameters and target appearance feature parameters to form cross-frame target association results. Specifically, the calculation of target matching values ​​based on target location parameters and target appearance feature parameters involves: For each two adjacent video frames, read the target candidate set. For each target candidate in the previous frame and each target candidate in the next frame, construct a candidate pair. Calculate the position difference based on the target position parameters of the two target candidates and normalize it to generate position similarity. Calculate the appearance difference based on the target appearance feature parameters and normalize it to generate appearance similarity. Sum the position similarity and appearance similarity with weights to obtain the target matching value of the candidate pair. Based on the cross-frame target association results, a unified behavior object identifier is assigned to the target candidates on the same association link, and the behavior object corresponding to the behavior object identifier is determined at each time sampling point of the video frame sequence, forming a time sampling point-level behavior object sequence. For each behavior object identifier, the target position parameters corresponding to the behavior object at each time sampling point are aggregated to construct a target trajectory sequence of the behavior object at continuous time sampling points. Specifically, constructing the target trajectory sequence of the behavior object at continuous time sampling points involves: For each behavior object identifier, the target position parameters corresponding to the behavior object at each time sampling point are read and arranged and connected in order. When the time sampling point has the target position parameters corresponding to the behavior object, the target position parameters are directly added to the target trajectory sequence in order. When the time sampling point is missing the target position parameters corresponding to the behavior object, the empty position mark corresponding to the time sampling point is recorded in the target trajectory sequence. All time sampling points are traversed in turn to form the target trajectory sequence of the behavior object.

[0022] In this embodiment, the formation of the behavioral feature time series includes: Based on the target trajectory sequence, for each action object, the corresponding target position parameters are read at the current time and the previous time sampling point. The displacement between adjacent time sampling points is calculated, and the velocity, acceleration, and motion direction angle are calculated based on the displacement and a fixed time interval. The calculation of the displacement between adjacent time sampling points specifically involves: Read the target position parameters corresponding to the current and previous time sampling points respectively, obtain the horizontal and vertical coordinates of the center point of the target box at the two time points, take the difference of the center point coordinates at the two time points as the displacement component, and obtain the horizontal displacement component and the vertical displacement component. The combined displacement length obtained by taking the square root of the sum of the squares of the horizontal displacement component and the vertical displacement component is the displacement amount. The calculation of velocity, acceleration, and motion direction angle based on displacement and a fixed time interval is as follows: The velocity is obtained by dividing the displacement by a fixed time interval, and the acceleration corresponding to the current time sampling point is obtained by dividing the difference between the velocities corresponding to two adjacent time intervals by a fixed time interval. The motion direction angle is obtained by taking the arctangent of the ratio of the longitudinal displacement component to the lateral displacement component. The change in direction is calculated based on the motion direction angle of adjacent time sampling points, and the trajectory curvature feature is calculated based on the change in direction and displacement. Simultaneously, the presence of a stationary state is determined based on the velocity and velocity threshold. The number of time sampling points that continuously satisfy the stationary state is accumulated to obtain the stationary duration feature, where: The calculation of the change in direction is as follows: The difference between the current motion direction angle and the previous time sampling point is used to obtain the direction angle change value. When the direction angle change value is negative, the absolute value is taken as the direction change amount. When the direction angle change value is greater than 180°, the direction angle change value is subtracted from 360° as the direction change amount. When the direction angle change value is positive and less than 180°, it is directly used as the direction change amount. The calculation of the trajectory curvature feature is specifically as follows: For the current time sampling point, the ratio between the change in direction and the displacement is used as the trajectory curvature feature. When the displacement is zero, the trajectory curvature feature is recorded as zero. The obtained residence time feature is specifically as follows: For each time sampling point, when the velocity is less than the velocity threshold, the time sampling point is determined to be in a stationary state. The time sampling points that are continuously determined to be in a stationary state are counted sequentially, and the count value is multiplied by a fixed time interval to obtain the stationary duration feature of the behavior object in the current time period. When the velocity of the time sampling point is not less than the velocity threshold, the continuous count value is cleared to zero and the accumulation starts again. The displacement, velocity, acceleration, motion direction angle, change of direction, trajectory curvature features, and dwell time features obtained at the current time sampling point are combined sequentially to form the behavior feature vector corresponding to the current time sampling point, thus obtaining the behavior feature time series. Specifically, obtaining the behavior feature time series involves: For each behavior object, according to the order of time sampling points, the displacement, velocity, acceleration, motion direction angle, direction change, trajectory curvature features and dwell time features corresponding to the current time sampling point are read in sequence and concatenated to form a behavior feature vector. The feature concatenation is repeated for all time sampling points to form a behavior feature time series.

[0023] In this embodiment, generating the corresponding behavioral response output sequence includes: An improved CMAC network is constructed. For each time sampling point in the behavioral feature time series, the behavioral features of the current time sampling point are divided according to the granularity level of the behavioral features, resulting in a multi-granularity feature set. Specifically, the division of the behavioral features of the current time sampling point according to the granularity level of the behavioral features is as follows: Based on the behavioral feature vector, each behavioral feature is assigned to different granularity levels according to the behavioral scale difference of the behavioral features. Displacement, velocity, acceleration and motion direction angle are regarded as fine-grained features, direction change and trajectory curvature features are regarded as medium-grained features, and dwell time features are regarded as coarse-grained features, thus obtaining the multi-granularity feature set corresponding to the current time sampling point. For each granularity level in the multi-granularity feature set, input space partitioning rules and quantization unit sets corresponding to each granularity level are configured. Furthermore, quantization interval numbering rules and quantization address encoding rules are configured for each granularity level, forming an input space quantization configuration that corresponds one-to-one with each granularity level. The specific configuration of the input space partitioning rules and quantization unit sets corresponding to the granularity levels is as follows: Based on the behavioral feature items contained in each granularity level, and taking the value range of each behavioral feature item as the input space, the value range is divided into 16 intervals by continuous division according to a fixed interval length. The value range of each behavioral feature item is divided into several non-overlapping quantization intervals. Each quantization interval corresponds to a quantization unit. The quantization unit represents the state of the behavioral feature value falling into the quantization interval. The interval division and quantization unit setting operations are performed on all behavioral feature items in the same granularity level to form the input space division rules and quantization unit set corresponding to the granularity level. The specific steps involve configuring quantization interval numbering rules and quantization address encoding rules for each granularity level as follows: For all quantization intervals within the same granularity level, unique interval numbers are assigned sequentially according to the order of the quantization intervals in the input space. The quantization interval number corresponding to the value of each behavioral feature at the current time sampling point is used as the quantization index of the behavioral feature. The quantization indices corresponding to each behavioral feature are combined in sequence to form a quantization address code. For each time sampling point, based on the input space quantization configuration, input space quantization mapping is performed on each feature at each granularity level to obtain a set of quantization interval numbers corresponding to each component of the granularity level feature. A quantization address code is then generated based on the quantization interval number set, and the center index position corresponding to the quantization address code is determined. Where: The specific steps for performing input space quantization mapping on each feature at each granularity level are as follows: For the current time sampling point, read the behavioral feature subsets corresponding to each granularity level, locate the value of each behavioral feature in the corresponding input space, map the current value of the behavioral feature to the quantization interval it falls into, and read the interval number corresponding to the quantization interval as the quantization result of the behavioral feature at the current time sampling point. The process of generating quantized address codes based on the quantized interval number set is as follows: According to the order of behavioral features within the granularity level, the quantization interval numbers corresponding to each behavioral feature are read sequentially, and the quantization interval numbers are combined in order to form a quantization address code; The determination of the center index position corresponding to the quantized address encoding specifically involves: Based on the arrangement position of the quantized address code in the quantized address space corresponding to the granularity level, the center index position corresponding one-to-one with the quantized address code is determined. The center index position represents the center number position of the quantized address code in the granularity level quantized address space. A memory index is generated based on the quantization address encoding corresponding to each granularity level. Within the memory cell set of each granularity level, a subset of activated memory cells is determined. Simultaneously, the relative positional relationship of the activated memory cell subsets in the quantization space is converted into structural offset information, which is recorded as relative offset labels. A position-sensitive structural memory matrix is ​​then constructed, where: The generation of the memory index based on the quantized address encoding corresponding to each granularity level is specifically as follows: Using the quantization address code as the index input, the index position of the quantization address code in the memory cell set is determined according to the correspondence between the quantization address code and the memory cell set in the granularity level, and recorded as the memory index corresponding to the current time sampling point in the granularity level; A memory unit set refers to a set of parameter storage units with fixed index numbers set for each granularity level. Each memory unit corresponds to a specific quantization address location in the input space quantization configuration and structurally contains the parameter storage location corresponding to the index number. The process of determining the subset of activated memory units within the set of memory units at each granularity level specifically involves: For each granularity level, the memory unit corresponding to the memory index is taken as the central memory unit, and multiple adjacent memory units within the index neighborhood of the central memory unit are selected as activated memory units to form a subset of activated memory units corresponding to the current time sampling point at the granularity level. The construction of the position-sensitive structural memory matrix is ​​specifically as follows: Using the memory index of the central memory unit as a reference, the index difference between the memory index of each activated memory unit and the central memory index is calculated as the relative offset of the corresponding activated memory unit. The relative offsets are combined according to the order of the activated memory units in the index space to form structural offset information. The structural offset information is then organized and stored in matrix form to construct a position-sensitive structural memory matrix. Based on a position-sensitive structural memory matrix, the weight storage structure is set as a tensor-type parameter structure. Tensorized weight indexing is performed on the tensor-type parameter structure based on the memory index to obtain the intermediate response output corresponding to each granularity level, where: The specific steps of setting the weight storage structure as a tensor parameter structure are as follows: Using the granular hierarchical structure, memory unit hierarchical structure, and relative structural positional relationships between memory units defined in the position-sensitive structural memory matrix as the basis for parameter organization, a multi-dimensional weighted storage space is constructed. The first dimension distinguishes different granular levels, the second dimension distinguishes different memory unit hierarchies within the same granular level, and the third dimension distinguishes the relative positional offset of memory units in the structural memory matrix. The one-dimensional weighted storage method, which uses a single index number for addressing, is mapped to a parameter storage method, which uses multi-dimensional index coordinates for addressing, thus forming a tensor-type parameter structure. The process of performing tensor quantization weight indexing on tensor parameter structures based on memory indexing is as follows: For the current time sampling point and the current granularity level, based on the determined memory index, the corresponding memory unit level number and structural offset information, construct an index coordinate combination that is consistent with the dimensional order of the tensor parameter structure, locate the corresponding parameter position in the tensor parameter structure, and read the corresponding parameter value as the intermediate response output of the granularity level at the current time sampling point. Position-aware attenuation weights are generated based on the relative distance between the features at each granularity level and the corresponding center index position. The intermediate response outputs corresponding to each granularity level are then weighted and synthesized to obtain the behavioral response output values ​​corresponding to the time sampling points, generating a behavioral response output sequence, where: The generation of the location-aware attenuation weights is specifically as follows: For the current time sampling point and each granularity level, the memory index corresponding to each activated memory unit in the granularity level is read, and the center index position corresponding to the granularity level is used as a reference. The index difference between each memory index position and the center index position is calculated as the relative distance. The weight corresponding to the memory index position with a relative distance of zero is set to the maximum value, and the weight value is decreased sequentially as the relative distance increases. A position-aware attenuation weight corresponding to the relative distance is generated for each activated memory unit. The generated behavioral response output sequence is specifically as follows: For the current time sampling point, each intermediate response output within the same granularity level is multiplied with its corresponding attenuation weight, and the multiplication results are summed to obtain the weighted response value corresponding to the granularity level at the current time sampling point. The weighted response values ​​corresponding to each granularity level are arranged and combined according to the fixed order of the granularity level to form the behavioral response output value corresponding to the current time sampling point. The weighted calculation and sequential combination process is repeated for consecutive time sampling points. The behavioral response output values ​​corresponding to each time sampling point are stored in chronological order to form a behavioral response output sequence.

[0024] In this embodiment, the combination to form the hypothesis parameter generation sequence includes: Based on the temporal order of the time sampling points, establish corresponding time sampling point identifiers for each behavioral response output value in the behavioral response output sequence, forming a behavioral response output index sequence aligned with time sampling points; In the behavior response output index sequence, a normal reference period and an abnormal reference period are determined, and the set of behavior response output values ​​within the normal reference period is denoted as the normal reference set, and the set of behavior response output values ​​within the abnormal reference period is denoted as the abnormal reference set. Specifically, determining the normal reference period and the abnormal reference period in the behavior response output index sequence involves: In the behavior response output index sequence, the time sampling point is used as the sequential index to sequentially traverse the behavior response output values ​​corresponding to consecutive time sampling points. The anomaly judgment threshold is used as the basis for dividing the time period. When the behavior response output value in a consecutive time period is less than the anomaly judgment threshold at all corresponding time sampling points, the consecutive time period is determined as a normal reference period. When the behavior response output value in a consecutive time period is greater than the anomaly judgment threshold at all corresponding time sampling points, the consecutive time period is determined as an abnormal reference period. Based on the normal reference set, the corresponding statistical parameter entries are calculated in order of time sampling points, and then arranged in order of time sampling points to form a sequence of normal behavior hypothesis parameters. Each statistical parameter entry includes the mean and variance parameters of the corresponding time sampling point. The calculation of the corresponding statistical parameter entries in order of time sampling points is as follows: In the normal reference set, for each time sampling point, the corresponding behavioral response output value and all behavioral response output values ​​within the corresponding sliding statistical window are read. The sum of all behavioral response output values ​​within the sliding statistical window is divided by the number of behavioral response output values ​​within the corresponding window to obtain the mean parameter corresponding to the time sampling point. Then, the difference between each behavioral response output value and the mean parameter is calculated. The differences are squared and summed. The sum of squares is divided by the number of behavioral response output values ​​to obtain the variance parameter corresponding to the time sampling point. Based on the anomaly reference set, the corresponding statistical parameter entries are calculated in order of time sampling points, and then arranged in order of time sampling points to form a sequence of hypothetical parameters for anomaly behavior. Each statistical parameter entry includes the mean and variance parameters for the corresponding time sampling point. The calculation of the corresponding statistical parameter entries in order of time sampling points is specifically as follows: In the abnormal reference set, for each time sampling point, the corresponding behavioral response output value and all behavioral response output values ​​within the corresponding sliding statistical window are read. The sum of all behavioral response output values ​​within the sliding statistical window is divided by the number of behavioral response output values ​​within the corresponding window to obtain the mean parameter corresponding to the time sampling point. Then, the difference between each behavioral response output value and the mean parameter is calculated. The differences are squared and summed. The sum of squares is divided by the number of behavioral response output values ​​to obtain the variance parameter corresponding to the time sampling point. The normal behavior hypothesis parameter sequence and the abnormal behavior hypothesis parameter sequence are combined in the order of time sampling points to obtain the hypothesis parameter generation sequence.

[0025] In this embodiment, the formation of the final cumulative statistic includes: Generate a sequence based on the hypothesis parameters, and read the corresponding behavioral response output value and the corresponding normal behavior hypothesis parameters and abnormal behavior hypothesis parameters for each time sampling point; For each time sampling point, based on the normal behavior hypothesis parameters and the abnormal behavior hypothesis parameters, the probability density values ​​of the current behavior response output value under the normal hypothesis period and the abnormal hypothesis period are calculated respectively. The ratio of the abnormal hypothesis probability density value to the normal hypothesis probability density value is used as the likelihood ratio of the time sampling point. The logarithm of the likelihood ratio is taken to obtain the log-likelihood increment. The specific steps for calculating the probability density values ​​of the current behavior response output value under the normal hypothesis period and the abnormal hypothesis period are as follows: For the current sampling point, under the normal assumption period, the difference between the current behavior response output value and the mean parameter in the normal behavior assumption parameters is calculated, and combined with the corresponding variance parameter, the probability of the difference occurring under the normal behavior assumption distribution is numerically calculated to obtain the probability density value of the current behavior response output value under the normal assumption period. Under the abnormal assumption period, the difference between the current behavior response output value and the mean parameter in the abnormal behavior assumption parameters is calculated, and combined with the corresponding variance parameter, the probability of the difference occurring under the abnormal behavior assumption distribution is numerically calculated to obtain the probability density value of the current behavior response output value under the abnormal assumption period. Sequential accumulation is performed based on the log-likelihood increment, and a directional gain coefficient is set according to the sign relationship between the log-likelihood increments of the current time sampling point and the previous time sampling point. After weighting the log-likelihood increment of the current time sampling point, the initial cumulative statistic is obtained, where: The setting of the directional gain coefficient is specifically as follows: For the current time sampling point, if the log-likelihood increment of the current time sampling point has the same sign as the log-likelihood increment of the previous time sampling point, the directional gain coefficient is set to a fixed enhancement coefficient greater than unity gain; otherwise, the directional gain coefficient is set to a fixed suppression coefficient less than unity gain. The initial cumulative statistic is obtained as follows: The log-likelihood increment corresponding to the current time sampling point is weighted with the directional gain coefficient to obtain the weighted log-likelihood increment of the current time sampling point. The weighted log-likelihood increment is then added to the cumulative statistic corresponding to the previous time sampling point. The cumulative statistic is set to 0 at the first time sampling point to obtain the initial cumulative statistic corresponding to the current time sampling point. A threshold adjustment factor is generated based on the variation amplitude of the behavioral response output values ​​between adjacent time sampling points and the probability density value of the normal hypothesis. The basic upper and lower bound thresholds are then updated according to this threshold adjustment factor to obtain the dynamic upper and lower bound thresholds corresponding to the current time sampling point, where: The generation of the threshold adjustment factor is specifically as follows: Read the behavioral response output values ​​corresponding to the current time sampling point and the previous time sampling point, calculate the difference between the two and take the absolute value as the change amplitude, combine the change amplitude with the normal hypothesis probability density value, perform weighted summation and normalization operation to obtain the threshold adjustment factor; The process of obtaining the dynamic upper bound threshold and dynamic lower bound threshold corresponding to the current time sampling point is as follows: Read the basic upper bound threshold and the basic lower bound threshold, perform a weighted numerical correction on the basic upper bound threshold based on the threshold adjustment factor to obtain the dynamic upper bound threshold, and perform a weighted numerical correction on the basic lower bound threshold based on the difference between 1 and the threshold adjustment factor to obtain the dynamic lower bound threshold. An anomaly pattern library is constructed. For the current time sampling point, based on the matching results between the behavior response output subsequence containing the current time sampling point and each anomaly pattern template, a pattern matching probability is generated, where: The construction of the exception mode library specifically involves: Based on the labeled abnormal behavior samples in historical surveillance video data, the corresponding behavior response output time series segments are extracted and organized according to the morphological characteristics of behavior response changes over time to form abnormal pattern templates. Each abnormal pattern template consists of behavior response output values ​​arranged in chronological order and corresponds to an abnormal change pattern. All abnormal pattern templates together constitute an abnormal pattern library. The response output subsequence refers to a continuous time segment sequence formed by selecting the behavior response output value corresponding to the current time sampling point in the behavior response output sequence in chronological order, with the current time sampling point as the center. The generated pattern matching probability is specifically as follows: For the current time sampling point, the similarity between the behavior response output subsequence and each abnormal pattern template in the abnormal pattern library is calculated, and the similarity results corresponding to each abnormal pattern template are normalized and mapped to a value between zero and one, which is used as the pattern matching probability corresponding to the current time sampling point. The update magnitude of the initial cumulative statistic corresponding to the current time sampling point is adjusted based on the pattern matching probability. Sequential accumulation is performed on the log-likelihood increment based on two different time scales to form a multi-time-scale cumulative statistical trajectory. The initial cumulative statistic is then weighted and fused based on the consistency results among the cumulative statistical trajectories of each time scale to obtain the final cumulative statistic, where: The adjustment of the initial cumulative statistics update magnitude corresponding to the current time sampling point based on pattern matching probability is as follows: At the current time sampling point, the corresponding pattern matching probability is read, and the update increment of the initial cumulative statistics is scaled according to the magnitude of the pattern matching probability. The update increment of the initial cumulative statistics corresponding to the time sampling point where the pattern matching probability is greater than the probability threshold is amplified, while the update increment of the initial cumulative statistics corresponding to the time sampling point where the pattern matching probability is less than the probability threshold is reduced, thus obtaining the update result of the initial cumulative statistics after pattern matching probability adjustment. The formation of multi-timescale cumulative statistical trajectories specifically involves: Sequential accumulation processing is performed on the current time sampling point and its corresponding log-likelihood increment at two different time scales. The log-likelihood increment corresponding to the current time sampling point is superimposed and updated with the cumulative statistics of the previous time sampling point. Each time scale corresponds to an independent time sampling span. At a short time scale, a rapidly changing cumulative statistical trajectory is formed with adjacent time sampling points as continuous update units. At a long time scale, a smoothly changing cumulative statistical trajectory is formed with 3 time sampling points as one cumulative update unit, thus forming a multi-time-scale cumulative statistical trajectory.

[0026] The final cumulative statistic is obtained as follows: For the current time sampling point, the statistical values ​​corresponding to the cumulative statistical track at each time scale are read, and the consistency of the statistical values ​​at each time scale is judged. The direction and relative magnitude of the change of the cumulative statistical values ​​at different time scales are compared. When the cumulative statistical values ​​at each time scale show the same direction of change at the same time sampling point, and the magnitude of the change is within the allowable deviation range, they are judged to be consistent. The statistical values ​​at each time scale are weighted and fused to obtain the final cumulative statistics. In the sequential accumulation process, the behavioral response entropy value is calculated based on the probability density values ​​of the normal hypothesis and the abnormal hypothesis, and the behavioral response entropy value is accumulated. When the accumulated behavioral response entropy value reaches the entropy threshold, the final accumulated statistic is reset; otherwise, the final accumulated statistic, the dynamic upper bound threshold, the dynamic lower bound threshold, and the early judgment flag are output, where: The calculation of the behavioral response entropy value is as follows: Read the normal hypothesis probability density value and the abnormal hypothesis probability density value corresponding to the time sampling point respectively, and normalize the two to form the probability distribution corresponding to the current time sampling point. Sum the information content corresponding to the normal hypothesis probability density value and the abnormal hypothesis probability density value to obtain the behavior response entropy value. The reset of the final cumulative statistics is specifically performed as follows: When the cumulative behavioral response entropy value corresponding to the current time sampling point reaches the entropy threshold, the final cumulative statistic corresponding to the current time sampling point is set to the initial statistical state, and the statistical state accumulated before the corresponding time sampling point is cleared at the same time. Then, the log-likelihood increment calculation, sequential accumulation and threshold determination process are performed sequentially for the time sampling points after the current time sampling point. The advance judgment flag is determined by the cumulative result of the behavioral response entropy value. When the cumulative behavioral response entropy value has not reached the entropy threshold, the advance judgment flag is in a valid state; otherwise, the advance judgment flag is in an invalid state.

[0027] In this embodiment, the output of abnormal behavior detection results includes: At each time sampling point, the final cumulative statistic is compared with the dynamic upper bound threshold and the dynamic lower bound threshold. If the pre-judgment flag is valid, an anomaly candidate result is directly generated; if the pre-judgment flag is invalid, an anomaly candidate result is generated based on the comparison result. The step of directly generating anomaly detection candidate results when the pre-determination flag is valid is as follows: When the early judgment flag is in a valid state, when the final cumulative statistic exceeds the dynamic upper limit threshold, the anomaly judgment candidate result corresponding to the current time sampling point is marked as an abnormal state; when the final cumulative statistic is lower than the dynamic lower limit threshold, the anomaly judgment candidate result corresponding to the current time sampling point is marked as a normal state, and the marking result is used as the anomaly judgment candidate result of the current time sampling point. The step of generating anomaly detection candidate results based on the comparison results when the pre-determination flag is invalid is as follows: When the early judgment flag is invalid, when the final cumulative statistic exceeds the dynamic upper limit threshold, the abnormal judgment candidate result is marked as abnormal; when the final cumulative statistic is lower than the dynamic lower limit threshold, the abnormal judgment candidate result is marked as normal; when the final cumulative statistic is between the dynamic upper limit threshold and the dynamic lower limit threshold, the abnormal judgment candidate result is marked as pending judgment. For the pending states in the anomaly detection candidate results, a consistency judgment value is generated based on the consistency results of pattern matching probability and multi-timescale cumulative statistical trajectory. Confirmation and cancellation operations are then performed on the anomaly detection candidate results to obtain the final anomaly detection result corresponding to the time sampling point, where: The consistency determination value is generated based on the consistency results of pattern matching probability and multi-timescale cumulative statistical trajectory, specifically as follows: For the current time sampling point, when the candidate result of the anomaly judgment is in the pending judgment state, the consistency result of the corresponding pattern matching probability and the cumulative statistical trajectory of multiple time scales at the time sampling point is read and numerical combination calculation is performed to generate the consistency judgment value corresponding to the current time sampling point. The specific steps for performing confirmation and cancellation operations on candidate results of anomaly determination are as follows: When the consistency judgment value exceeds the consistency judgment threshold, the abnormal judgment candidate result in the pending judgment state is confirmed and marked as abnormal. When the consistency judgment value does not exceed the consistency judgment threshold, the abnormal judgment candidate result in the pending judgment state is withdrawn and marked as normal, thus obtaining the final abnormal judgment result corresponding to the current time sampling point. When the final anomaly determination result is an anomaly, the anomaly behavior detection result corresponding to the time sampling point is output, where: Abnormal behavior detection results refer to the structured detection output information generated corresponding to the time sampling point when the final abnormal judgment result corresponding to the time sampling point is marked as an abnormal state. This includes the time identifier of the time sampling point, the corresponding behavior object identifier, and the abnormal state identifier.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a comprehensive transportation hub in a city center. This hub has an average daily passenger flow of approximately 180,000 people, and the monitoring area includes key areas such as the waiting hall, security checkpoints, and entrance / exit channels. Previous anomaly detection methods based on fixed thresholds and single-frame judgment were prone to false alarms during peak hours due to fluctuations in passenger flow, and often exhibited delayed responses in the early stages of abnormal behavior, failing to meet the timeliness and stability requirements of on-site management.

[0029] In this scenario, the method of this invention is deployed in an existing monitoring system. Continuous video captured by monitoring cameras is sampled at a uniform time interval of 1 second, forming a continuous sequence of video frames. Pedestrians in the video are detected and cross-frame consistency is correlated to construct the continuous motion trajectory of each pedestrian within the monitored area. Behavioral features such as displacement, speed, direction change, and dwell time are extracted from the trajectory to form a behavioral feature time series. An improved CMAC network is used to quantize and map the behavioral features, generating a stable behavioral response output sequence. The behavioral responses are statistically accumulated second by second based on a sequential probability ratio test, and combined with anomaly pattern matching and multi-timescale consistency judgment, anomaly judgment results are continuously output.

[0030] In a 12-week continuous operational test, approximately 425 hours of valid surveillance video data were collected, covering various typical time periods including weekdays, weekends, and the periods before and after holidays. Manual annotation results showed 104 instances of clearly abnormal behavior during this period. The method of this invention detected 101 abnormal behaviors, with 3 missed detections, achieving an overall detection rate of 97.1%. The number of false alarms was 7, mainly concentrated during peak periods with high pedestrian traffic and significant short-term fluctuations, resulting in a false alarm rate of approximately 6.7%. In comparison, the existing fixed threshold method detected 92 abnormal behaviors, with 12 missed detections and 40 false alarms, a significantly higher false alarm rate than the method of this invention.

[0031] As can be seen from the embodiments, the present invention can effectively characterize the evolution of behavior over time in complex crowd environments, avoiding false alarms caused by short-term fluctuations.

[0032] Table 1. Comparison of detection effects between the method of the present invention and the traditional method under long-term operating conditions.

[0033] As can be seen from the data in Table 1, during the 12-week continuous monitoring of abnormal behavior in surveillance videos, the method of this invention and the traditional fixed threshold method showed a significant difference in overall detection performance. In terms of the total number of abnormal events, a total of 104 abnormal behavior events were manually identified and confirmed. The method of this invention detected 101 abnormal behaviors, with only 3 missed, while the traditional method detected 92, with 12 missed. This demonstrates that the present invention has a significant advantage in terms of overall coverage and completeness of abnormal behavior detection.

[0034] Analysis of false alarms showed that the method of this invention generated 7 false alarms within 12 weeks, mainly concentrated during peak periods of rapid changes in pedestrian density, with an overall false alarm rate of approximately 6.7%. In contrast, the traditional method generated 40 false alarms, a significantly higher rate. This indicates that the present invention, by introducing behavioral feature time-series modeling and sequential statistical cumulative judgment, effectively suppresses the interference of short-term fluctuations on the judgment results, and can significantly reduce the probability of false alarms while maintaining high sensitivity.

[0035] In terms of response efficiency, the method of this invention has an average judgment time of 7.9 seconds for abnormal behavior, while the average judgment time of traditional methods is 21.2 seconds. The method of this invention can provide stable statistical judgment results before the behavior is fully manifested, thus gaining more reaction time for on-site management and emergency response.

[0036] In summary, the data in Table 1 fully demonstrates that the method of the present invention is superior to traditional methods in key indicators such as abnormal behavior detection rate, false alarm control, and judgment timeliness. It can more realistically and continuously reflect the characteristics of behavior evolution over time, and verify the practicality and stability of the method of the present invention in complex monitoring scenarios.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting abnormal behavior of monitoring video data based on behavior feature recognition, characterized in that, include: Collect continuous monitoring video data and perform time sampling on the continuous monitoring video data according to a unified time base to form a video frame sequence; The video frame sequence is subjected to target detection and target consistency association processing to determine the behavioral object corresponding to each time sampling point and construct the corresponding target trajectory sequence. Behavioral features are extracted at each time sampling point to form a behavioral feature time series. Based on behavioral feature time series, an improved CMAC network is constructed to perform input space quantization mapping on the behavioral features at each time sampling point, activate the corresponding memory unit set, and generate the corresponding behavioral response output sequence. Based on the behavioral response output sequence, the corresponding statistical parameters of the behavioral response output sequence are extracted in the order of time sampling points in the normal and abnormal reference periods, and combined to form the hypothesis parameter generation sequence. Generate a sequence based on the hypothetical parameters, perform a sequential probability ratio test, construct a corresponding statistical observation for the behavioral response output value at each time sampling point, calculate the corresponding likelihood ratio and perform sequential cumulative processing to form the final cumulative statistic. Anomaly detection is performed based on the final cumulative statistics, and the abnormal behavior detection results are output.

2. The method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to claim 1, characterized in that, The process of forming a video frame sequence includes: Acquire continuous monitoring video data from surveillance camera equipment, record the time information corresponding to the continuous monitoring video data, and determine the start time point of video acquisition; Based on a unified time sampling rule set at the start time point, multiple time sampling points are generated on the time axis of continuous monitoring video data at fixed time intervals. For each time sampling point, the video frame closest to the time of the time sampling point is selected from the continuous monitoring video data and used as the video frame corresponding to the current time sampling point. The video frames corresponding to each time sampling point are arranged in chronological order to form a video frame sequence arranged in chronological order.

3. The method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to claim 1, characterized in that, The construction of the corresponding target trajectory sequence includes: Perform target detection processing on each video frame in the video frame sequence, obtain the target candidate set of all detected targets in the video frame, and record the target position parameters and target appearance feature parameters for each detected target in the video frame; Based on video frame sequences, target consistency association processing is performed on the target candidate sets in two adjacent video frames. The target matching value is calculated based on the target position parameters and target appearance feature parameters to form cross-frame target association results. Based on the cross-frame target association results, a unified behavior object identifier is assigned to the target candidates on the same association link, and the behavior object corresponding to the behavior object identifier is determined at each time sampling point of the video frame sequence, forming a time sampling point-level behavior object sequence. For each behavior object identifier, the target position parameters of the behavior object at each time sampling point are aggregated to construct the target trajectory sequence of the behavior object at continuous time sampling points.

4. The method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to claim 1, characterized in that, The formation of the behavioral feature time series includes: Based on the target trajectory sequence, for each action object, the corresponding target position parameters are read at the current time and the previous time sampling point, the displacement between adjacent time sampling points is calculated, and the velocity, acceleration and motion direction angle are calculated based on the displacement and a fixed time interval. The direction change is calculated based on the motion direction angle of adjacent time sampling points, and the trajectory curvature feature is calculated based on the direction change and displacement. At the same time, the presence of a stationary state is determined based on the velocity and velocity threshold. The number of time sampling points that continuously meet the stationary state is accumulated to obtain the stationary duration feature. The displacement, velocity, acceleration, motion direction angle, direction change, trajectory curvature features, and dwell time features obtained at the current time sampling point are combined in sequence to form the behavior feature vector corresponding to the current time sampling point, thus obtaining the behavior feature time series.

5. The method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to claim 1, characterized in that, The generation of the corresponding behavioral response output sequence includes: An improved CMAC network is constructed. For each time sampling point in the behavioral feature time series, the behavioral features of the current time sampling point are divided according to the granularity level of the behavioral features to obtain a multi-granularity feature set. For each granularity level in the multi-granularity feature set, the input space partitioning rules and quantization unit set corresponding to the granularity level are configured respectively, and the quantization interval numbering rules and quantization address encoding rules are configured for each granularity level respectively, forming an input space quantization configuration that corresponds one-to-one with each granularity level; For each time sampling point, based on the input space quantization configuration, the input space quantization mapping is performed one by one for each granularity level feature to obtain the quantization interval number set corresponding to each component of the granularity level feature, and the quantization address code is generated based on the quantization interval number set, while the center index position corresponding to the quantization address code is determined. Memory indexes are generated based on the quantization address encoding corresponding to each granularity level, and the activated memory cell subsets are determined in the memory cell set of each granularity level. At the same time, the relative positional relationship of the activated memory cell subsets in the quantization space is converted into structural offset information, and the structural offset information is recorded as relative offset labels to construct a position-sensitive structural memory matrix. Based on the position-sensitive structural memory matrix, the weight storage structure is set as a tensor parameter structure. Tensorized weight indexing is performed on the tensor parameter structure based on the memory index to obtain the intermediate response output corresponding to each granularity level. Position-aware attenuation weights are generated based on the relative distance between the features of each granular level and the corresponding center index position. The intermediate response outputs corresponding to each granular level are weighted and synthesized to obtain the behavioral response output values ​​corresponding to the time sampling points, and a behavioral response output sequence is generated.

6. The method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to claim 1, characterized in that, The combination forms the hypothesis parameter generation sequence, including: Based on the temporal order of the time sampling points, establish corresponding time sampling point identifiers for each behavioral response output value in the behavioral response output sequence, forming a behavioral response output index sequence aligned with time sampling points; In the behavior response output index sequence, determine the normal reference period and the abnormal reference period, and denote the set of behavior response output values ​​within the normal reference period as the normal reference set, and denote the set of behavior response output values ​​within the abnormal reference period as the abnormal reference set. Based on the normal reference set, the corresponding statistical parameter entries are calculated according to the time sampling point order, and the statistical parameter entries are arranged according to the time sampling point order to form a normal behavior hypothesis parameter sequence. Each statistical parameter entry includes the mean parameter and variance parameter of the corresponding time sampling point. Based on the abnormal reference set, the corresponding statistical parameter entries are calculated according to the time sampling point order, and the statistical parameter entries are arranged according to the time sampling point order to form the abnormal behavior hypothesis parameter sequence. Each statistical parameter entry includes the mean parameter and variance parameter of the corresponding time sampling point. The normal behavior hypothesis parameter sequence and the abnormal behavior hypothesis parameter sequence are combined in the order of time sampling points to obtain the hypothesis parameter generation sequence.

7. The method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to claim 1, characterized in that, The formation of the final cumulative statistic includes: Generate a sequence based on the hypothesis parameters, and read the corresponding behavioral response output value and the corresponding normal behavior hypothesis parameters and abnormal behavior hypothesis parameters for each time sampling point; For each time sampling point, based on the normal behavior hypothesis parameters and the abnormal behavior hypothesis parameters, the probability density values ​​of the current behavior response output value under the normal hypothesis period and the abnormal hypothesis period are calculated respectively. The ratio of the abnormal hypothesis probability density value to the normal hypothesis probability density value is used as the likelihood ratio of the time sampling point. The logarithm of the likelihood ratio is taken to obtain the log-likelihood increment. Sequential accumulation is performed based on the log-likelihood increment, and a directional gain coefficient is set according to the sign relationship between the log-likelihood increment of the current time sampling point and the previous time sampling point. After weighting the log-likelihood increment of the current time sampling point, the initial cumulative statistic is obtained. A threshold adjustment factor is generated based on the variation amplitude between the behavioral response output values ​​of adjacent time sampling points and the normal hypothesis probability density value. The basic upper limit threshold and the basic lower limit threshold are updated according to the threshold adjustment factor to obtain the dynamic upper limit threshold and the dynamic lower limit threshold corresponding to the current time sampling point. Construct an abnormal pattern library. For the current time sampling point, generate the pattern matching probability based on the matching results between the behavior response output subsequence containing the current time sampling point and each abnormal pattern template. The update magnitude of the initial cumulative statistics corresponding to the current time sampling point is adjusted based on the pattern matching probability. Sequential accumulation is performed on the log-likelihood increment based on two different time scales to form a multi-time-scale cumulative statistical trajectory. The initial cumulative statistics are weighted and fused based on the consistency results between the cumulative statistical trajectories of each time scale to obtain the final cumulative statistics. During the sequential accumulation process, the behavioral response entropy value is calculated based on the probability density values ​​of the normal hypothesis and the probability density values ​​of the abnormal hypothesis, and the behavioral response entropy value is accumulated. When the accumulated behavioral response entropy value reaches the entropy threshold, the final accumulated statistic is reset; otherwise, the final accumulated statistic, the dynamic upper bound threshold, the dynamic lower bound threshold, and the early judgment flag are output.

8. The method for detecting abnormal behavior in surveillance video data based on behavioral feature recognition according to claim 1, characterized in that, The output abnormal behavior detection results include: At each sampling point, the final cumulative statistic is compared with the dynamic upper bound threshold and the dynamic lower bound threshold. When the advance judgment flag is valid, an anomaly judgment candidate result is directly generated. When the advance judgment flag is invalid, an anomaly judgment candidate result is generated based on the comparison result. For the pending states in the anomaly detection candidate results, a consistency judgment value is generated based on the consistency results of pattern matching probability and multi-timescale cumulative statistical trajectory. Confirmation and cancellation operations are performed on the anomaly detection candidate results to obtain the final anomaly detection result corresponding to the time sampling point. When the final anomaly determination result is an anomaly, the abnormal behavior detection result corresponding to the time sampling point is output.

Images