Reconnaissance behavior big data identification system and method for smart port

Abstract

The invention discloses an investigation behavior big data identification system and method for an intelligent port, and belongs to the technical field of computer vision and unmanned aerial vehicle monitoring. The system comprises an unmanned aerial vehicle cluster, an edge computing node and a cloud management platform. The video time sequence information density is improved through a linear frame insertion module; the high-order feature coding module extracts static and dynamic angle features including an adjacent node angle, a center orientation angle, a symmetric node angle and a change rate thereof, and constructs a fusion feature matrix; the lightweight graph convolutional network is dynamically divided and grouped based on joint point motion amplitude, and key region features are enhanced in combination with a space grouping attention mechanism. According to the method, through dynamic grouping adaptive region division and time sequence feature fusion, the problems of coverage blind areas, noise interference and insufficient behavior recognition precision in the prior art are solved. The method has the effects that the parameter quantity is reduced by 33%, and adaptive edge calculation is performed; and the accuracy of Top-1 is improved by 4.15%.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer vision, drone monitoring and big data analysis, and specifically to a big data identification system and method for investigative behavior used in smart ports. Background Art

[0002] Traditional smart port monitoring relies on fixed cameras and manual inspections, and has problems such as coverage blind spots, low data processing efficiency, and insufficient behavior recognition accuracy.

[0003] In the process of implementing the present invention, the inventors found that the following problems exist in the existing technology: the deep learning model has high computing power requirements and is difficult to deploy on the drone platform; the feature representation is single and only relies on static angles, which cannot capture continuous changes in movements; the fixed grouping strategy is prone to missing key areas under noise interference, for example, wind blowing clothes or crowded scenes can lead to misjudgment, and it is difficult to distinguish smuggling gestures from normal movements.

[0004] The present invention fuses static and dynamic angle features through a high-order feature encoding module (HFBA) to enhance sensitivity to fast movements; dynamically divides and groups the image through a lightweight graph convolutional network (SGA-GCN) to adaptively focus on key areas, thereby overcoming the technical defects of noise interference, poor scene adaptability and high model complexity. Summary of the Invention

[0005] The purpose of the present invention is to provide a big data identification system and method for investigative behavior at smart ports to solve the problems raised in the above-mentioned background technology.

[0006] In order to solve the above technical problems, on the one hand, a big data identification system for investigative behavior for smart ports is provided, and on the other hand, a method for the big data identification system for investigative behavior for smart ports is provided.

[0007] A big data identification system for investigative behavior at smart ports, including:

[0008] The data acquisition and transmission module is used to collect aerial videos in real time through a drone cluster and transmit them to edge computing nodes through multi-node load balancing;

[0009] Linear interpolation module, used to dynamically insert intermediate frames into low-frame-rate videos to improve the density of temporal information;

[0010] High-order feature encoding module, used to extract the coordinates of human joints, calculate static angle features and dynamic angle change rate features, and generate a fusion feature matrix;

[0011] A lightweight graph convolutional network module is used to dynamically divide and group the motion amplitudes of the joints and enhance key area features in combination with a spatiotemporal attention mechanism;

[0012] The behavior classification and alarm module is used to trigger alarms based on behavior probabilities and record abnormal information.

[0013] Preferably, the static angle features extracted by the high-order feature encoding module include adjacent node angles, central orientation angles and symmetrical node angles, and the dynamic angle features are the rate of change of angles between adjacent frames, and noise is eliminated through Z-Score standardization and sliding average filtering.

[0014] Preferably, the SGA-GCN module predicts joint importance scores through a lightweight fully connected layer, dynamically merges the joint points to generate groups based on Euclidean distance, and enhances local action features through spatial average feature weighting.

[0015] A method for a big data identification system for investigative behavior at a smart port, comprising the following steps:

[0016] Aerial video is collected by drones, and the linear interpolation algorithm is used to unify the number of frames to the target length;

[0017] Extracting the coordinates of the human body joint points in each frame, and calculating the static angle feature and the dynamic angle change rate feature;

[0018] Dynamically dividing the groups based on the motion amplitude of the joint points, and enhancing the key area features through the spatial grouping attention mechanism;

[0019] The spatiotemporal features are integrated to classify behaviors, and an alarm is sent if the probability of abnormal behavior exceeds the threshold.

[0020] Preferably, the linear interpolation algorithm comprises the following steps:

[0021] S201: Calculate the number of frames that need to be inserted between adjacent frames: Number of frames to be deleted after insertion: f d =(f f +1)×f0-f u ; Among them, f u is the target frame number; f0 is the original frame number; ρ(x) is the rounding function;

[0022] S202: The joint point detection algorithm extracts N joint point coordinates in each frame;

[0023] S203: Calculate the displacement matrix between adjacent frames: M i =C i+1 -C i Among them, C i is the skeleton data extracted from the i-th frame of the original video;

[0024] S204: Calculate the single-step displacement matrix between adjacent frames: Pi =M i / f f ;

[0025] S205: Generate the k-th vertex λ of the i-th frame i Coordinate vector of: C i,k =C i +(k-1)×P i ;

[0026] Video data after interpolation F∈R C×T×N , where C is the coordinate dimension, T is the target frame length, i.e. 60, and N is the number of joint points.

[0027] Preferably, the dynamic angle change rate characteristic is expressed by the formula Calculate and combine with the static features to form a 6-dimensional fusion feature matrix; where R s,t (Ni) represents the static joint angle value of the i-th joint point and the t-th frame; Δt t Indicates the time interval between the t-th frame and the t-1-th frame.

[0028] Preferably, the dynamic grouping includes: normalizing the joint motion amplitude, generating the importance score through the Softmax function, screening the joint points with scores exceeding a threshold, and merging them into the dynamic grouping based on the Euclidean distance.

[0029] Preferably, the spatial grouping attention mechanism includes: calculating the spatial average features within the group, generating attention weights through the fully connected layer, weightedly enhancing the local action features and suppressing noise.

[0030] Preferably, the behavior classification adopts the joint optimization of cross entropy loss and angle feature regularization term, outputs the abnormal behavior probability distribution and sets 80% as the alarm threshold.

[0031] Preferably, the warning information includes the behavior type, timestamp and location coordinates, and is pushed to the law enforcement officer's terminal via SMS or platform interface.

[0032] One of the technical methods in the above technical solution has the following beneficial effects: due to the use of the fusion technology of static angle features and dynamic angle change rate features in the HFBA module, it overcomes the defects of a single static feature that cannot capture continuous changes in movements and insufficient sensitivity to rapid movements, thereby achieving high-precision recognition of subtle abnormal behaviors such as gestures for handing over contraband, and significantly enhancing the ability to distinguish similar movements.

[0033] Another technical method in the above technical solution has the following beneficial effects: Due to the adoption of the dynamic grouping mechanism and lightweight network structure design of the SGA-GCN module, it overcomes the problems of regional omissions, noise interference and large number of model parameters caused by fixed grouping, thereby achieving efficient deployment with a 33% reduction in parameters and a 24% shortening of training time, and is suitable for drone edge computing platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of the execution of the investigation behavior in the second embodiment of the present invention;

[0035] Figure 2 These are the joints of the human skeleton in the second embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0037] See also Figures 1 to 2 The present invention provides three embodiments. Embodiment 1 is a big data identification system for investigative behavior at smart ports, comprising: a data acquisition and transmission module, a linear interpolation (LI) module, a high-order feature encoding (HFBA) module, a lightweight graph convolutional network (SGA-GCN) module, a behavior classification and alarm module, and a cloud management platform.

[0038] The data collection and transmission module is applied to the local layer. It dynamically distributes the data of the drone cluster to different edge computing nodes through multi-node load balancing technology to avoid network congestion. It supports the real-time transport protocol RTMP to ensure low-latency data transmission.

[0039] LI module, applied to the local layer, is used for low frame rate video by calculating the inter-frame displacement matrix M i Dynamically insert intermediate frames to improve the density of temporal information;

[0040] The HFBA module is applied to the local layer to extract the coordinates of human joint points from the video, including adjacent joint angles, center orientation angles, and symmetric joint angles, and generate structured skeleton data;

[0041] SGA-GCN module, applied to the edge layer, lightweight graph convolutional network (SGA-GCN), consists of 7 basic blocks B1~B7, see Figure 2Each basic block contains a spatial GCN that analyzes the positional relationship between human joints, a spatiotemporal attention module (STC) that focuses on both the temporal changes and spatial positions of actions, a spatial grouping attention module (SGA) that divides the human body into multiple regions (such as the upper body and lower body) and focuses on enhancing the key parts of the current action, and a temporal GCN that analyzes the motion trajectory of joints over time for multi-branch feature fusion.

[0042] The behavior classification and alarm module, applied to the edge layer, stores the time, location, and category information of abnormal behavior. Based on the probability of the behavior, for example, if the probability of the behavior being classified as "smuggling" is greater than 80%, an alarm message is automatically triggered. The alarm information includes the behavior type, timestamp, and location coordinates, and is pushed to the law enforcement terminal via SMS or platform interface.

[0043] The cloud management platform, applied to the cloud layer, includes big data analysis and a visual dashboard. Big data analysis is used to aggregate data from multiple drones and discover behavioral patterns, such as areas with abnormally high incidence during peak hours. The visual dashboard is used to provide heat maps, alarm statistics, and real-time monitoring images to assist managers in decision-making.

[0044] The second embodiment is a method for a big data identification system for investigative behavior at a smart port, comprising the following steps:

[0045] Aerial video is collected by drones, and the linear interpolation algorithm is used to unify the number of frames to the target length;

[0046] Extracting the coordinates of the human body joint points in each frame, and calculating the static angle feature and the dynamic angle change rate feature;

[0047] Dynamically divide the groups based on the motion amplitude of the joint points, and enhance the key area features through the spatial grouping attention mechanism;

[0048] The spatiotemporal features are integrated to classify behaviors, and an alarm is sent if the probability of abnormal behavior exceeds the threshold.

[0049] Furthermore, the data preprocessing and linear interpolation (LI module) performs the following steps:

[0050] The number of frames in the original video captured by the drone may be inconsistent and needs to be padded to a fixed length. In the embodiment of the present invention, the fixed length is 60 frames. If the number of frames does not reach 60, an interpolation algorithm is used. If the number of frames exceeds 60, the video frames without "people" shots are removed. If the number of frames is insufficient after removal, the interpolation algorithm is used to fill the gap. The specific interpolation algorithm is as follows:

[0051] S201: Calculate the number of frames that need to be inserted between adjacent frames: f uis the target frame number (60), f0 is the original frame number, ρ(x) is the rounding function; the number of frames to be deleted after interpolation: f d =(f f +1)×f0-f u ;

[0052] S202: Joint point detection algorithm (OpenPose) extracts the coordinates of 18 joint points in each frame;

[0053] S203: Calculate the displacement matrix between adjacent frames: M i =C i+1 -C i , C i is the skeleton data extracted from the i-th frame of the original video;

[0054] S204: Calculate the single-step shift matrix between adjacent frames: P i =M i / f f ;

[0055] S205: Generate the k-th vertex λ of the i-th frame i Coordinate vector of: C i,k =C i +(k-1)×P i ; Video data after interpolation F∈R C×T×N ; Where C is the coordinate dimension, i.e. C=3, T is the number of frames, i.e. 60, and N is the number of joint points, i.e. 18.

[0056] Furthermore, the high-order bone angle feature encoding (HFBA module) performs the following steps:

[0057] Step 1: Input the interpolated video frame sequence (60fps) output by the LI module, including the skeletal joint coordinate data of each frame (18 joints, 3D coordinates), and add timestamp information: the acquisition time t of each frame (unit: milliseconds);

[0058] Step 2: Extract the skeleton data matrix C∈R from the interpolated video stream C×T×N , where C = 3, is the coordinate dimension (x, y, z), T = 18, is the number of joint points, and N = 60, is the number of frames;

[0059] Step 3: Extract timestamp vector Δt∈R T , where Δt i =t i -t i-1 , represents the time interval between adjacent frames;

[0060] Step 4: Define three high-order angles:

[0061] Step 4.1: Angles of adjacent nodes:

[0062] Step 4.1.1: Determine a target joint Ni. The computer defines it in sequence according to the joints included in the moving part.

[0063] Step 4.1.2: If Ni has one adjacent node joint, the angle is 0 and recorded as the symmetric node angle;

[0064] Step 4.1.3: If Ni has two adjacent node joints, the angle is calculated using the static joint angle calculation formula and recorded as the symmetrical node angle, and then proceed to Step 5;

[0065] Step 4.1.4: If Ni has more than two adjacent node joints, the angles are calculated using the static joint angle calculation formula and recorded as symmetrical node angles. The angle with the larger motion range is taken as the adjacent node angle, and then proceed to Step 5.

[0066] Step 4.2: Center orientation angle:

[0067] Step 4.2.1: Determine the target joint Ni. If Ni∈[N2,N4,N6,N8,N10,N12,N16], calculate the two joint angles of neck-Ni-left hip and neck-left hip-Ni and record them as the center orientation angles, then proceed to Step 5. N2 represents the second joint point, and so on.

[0068] Step 4.2.2: Determine the target joint Ni. If Ni∈[N3,N5,N7,N9,N11,N13,N17], calculate the two joint angles of neck-Ni-right hip and neck-right hip-Ni and record them as the center orientation angles, then proceed to Step 5.

[0069] Step 4.3: Symmetrical node angle: Determine the target joint Ni. If the angle between Ni and the symmetric joint is calculated, that is, [N2,...,N7,N10,...,N13,N16,N17], the angle is calculated using the static joint angle calculation formula and recorded as the symmetric node angle;

[0070] Table 1 Comparison table of joint points of human skeleton

[0071]

[0072]

[0073] Step 5: Static joint angle calculation formula:

[0074]

[0075] Among them, Ni is the target joint; a1, a2 are joint points; g, a1, a2 coordinates are (x Ni ,y Ni ,z Ni )、(x a1 ,y a1 ,z a1 )、(x a2 ,y a2 ,z a2 ), then the vector Ni pointing to nodes a1 and a2 is:

[0076]

[0077] Step 6: If the center orientation angle is not calculated, return to Step 4.2; if the symmetric node angle is not calculated, return to Step 4.3;

[0078] Step 7: Get the static angle feature matrix F of each joint containing three angle features static ∈R 3×60×18 ;

[0079] Step 8: For each angle feature of each joint, calculate the rate of change between adjacent frames:

[0080]

[0081] The change rate of the first frame is defined as ΔR s,1 =0; R s,t (Ni) represents the static joint angle value of the i-th joint point and the t-th frame; Δt t Indicates the time interval between the t-th frame and the t-1-th frame;

[0082] Step 9: Get the dynamic feature matrix F containing the three angle change rates of each joint dynamic ∈R 3 ×60×18 ;

[0083] Step 10: The static angle feature matrix F static ∈R 3×60×18 and the dynamic feature matrix F dynamic ∈R 3×60×18 Spliced ​​into fusion feature matrix F fused ∈R 6×60×18 ;

[0084] Step 11: Perform Z-Score normalization on the 6-dimensional features (3 static + 3 dynamic) of each joint to eliminate dimensional differences;

[0085] Step 12: Apply sliding average filtering (window size = 5 frames) to the dynamic features to reduce noise interference;

[0086] Step 13: Output the normalized fusion feature matrix F norm ∈R 6×60×18 , transmitted to the SGA-GCN module.

[0087] Furthermore, the Spatial Grouped Attention Graph Convolutional Network (SGA-GCN) performs the following steps:

[0088] M1: For each joint, calculate its range of motion: Where Nj is the joint point index;

[0089] M2: M Nj Perform maximum-minimum normalization and scale the values ​​to the [0,1] interval:

[0090]

[0091] M3: Set the network structure of the lightweight fully connected layer:

[0092] Input layer: 18 neurons, corresponding to the normalized motion amplitudes of 18 joints;

[0093] Output layer: 18 neurons, outputting the group importance score of each joint;

[0094] Activation function: Softmax, to ensure that the sum of the importance scores of all joints is 1;

[0095] M4: Input vector M norm,Nj Through the weight matrix W∈R 18×18 and the bias vector b∈R 18 Mapping to raw scores:

[0096] S raw =W×M norm,Nj +b

[0097] M5: Convert the original score into a probability distribution, that is, the normalized importance score of the Nj-th joint point, that is, the probability value:

[0098]

[0099] M6: Select score s Nj >τ joints, generate candidate list Nj candidate ={Nj|s Nj >τ}, where τ is the preset threshold, τ = 0.1;

[0100] M7: Extract the 3D coordinates of the candidate joint points of the current frame from the skeleton data matrix C;

[0101] M8: If the distance between two candidate joint points is ≤ dmax , then merge into the same group, where d max is the hierarchical clustering threshold based on Euclidean distance, d max =0.5m;

[0102] M9: Traverse all candidate joint points in Step 6 and execute M8. After the traversal, a dynamic group list G = {g1, g2, ..., gk} is generated. Each group contains the associated joint index.

[0103] M10: For each group gk, extract the feature subset F of its associated joint points gk ∈R 6×60×|gk| ;

[0104] M11: Calculate the spatial average features of each group gk:

[0105]

[0106] M12: Calculate the attention weight of each joint through the fully connected layer:

[0107]

[0108] M13: Weight the features of each joint point: F' Nj =F Nj w Nj ; Among them, feature weighting is used to enhance the motion characteristics of local joints and reduce the impact of noise;

[0109] M14: Output enhanced feature matrix F enhanced ∈R 6×60×18 , transmitted to the behavior classification module.

[0110] Furthermore, abnormal behavior classification and alarming are performed as follows:

[0111] Step 1: Perform average or maximum pooling on each type of feature (6 dimensions) in the time dimension (60 frames) to generate a global feature vector for each frame;

[0112] Step 2: Through the weight matrix W Nj ∈R 108×N and the bias vector b Nj ∈R N Map the features to the behavior category space to obtain the raw score z; where N is the number of behavior categories, such as "normal", "suspicious handover", "contraband carrying", etc.

[0113] Step 3: Convert the original score z into a probability distribution P∈R N ,For the behaviors classified as abnormal ones in the ,behavior category, they are marked as high-risk behaviors;

[0114] Step 4: If P 高风险行为 >δ, trigger an alarm, where δ is the threshold for triggering an alarm, δ = 0.8;

[0115] Step 5: Record the behavior type, timestamp, location coordinates (obtained from drone data), and probability value;

[0116] Step 6: Push to the law enforcement officer’s terminal via SMS / platform interface.

[0117] Example 3 is a specific implementation method of a big data identification system for investigative behavior at smart ports, including system deployment and hardware configuration, software framework and data flow, and examples of abnormal behavior identification;

[0118] The system deployment and hardware configuration include drone clusters, edge computing nodes and a cloud management platform; the drone cluster is equipped with high-resolution cameras (4K@60fps), infrared sensors and edge computing units (NVIDIA Jetson AGX Xavier) for real-time skeletal data extraction and compressed transmission; the edge computing nodes are equipped with NVIDIA RTX 4090 GPUs and 128GB of memory, deployed in the port monitoring center to run the SGA-GCN inference engine; the cloud management platform is built on a Kubernetes cluster, supports PB-level video data storage and distributed analysis, and provides visual dashboards such as behavior heat maps and alarm logs.

[0119] The software framework and data flow include data acquisition and transmission, preprocessing module implementation and SGA-GCN model training; data acquisition and transmission is that the drone transmits the video stream to the edge node through the RTMP protocol, the skeleton data is encapsulated in JSON format, and the frame rate is adaptively adjusted in the range of [15fps, 60fps]; the preprocessing module implementation uses OpenCV for frame extraction and implements the interpolation algorithm based on NumPy to output the tensor F∈R 3×60×18 ; Among them, 3 is the number of channels, 60 is the frame rate, and 18 is the number of joints. The SGA-GCN model training uses the UAV-Human extended version training dataset, which contains 200 types of behaviors in port scenarios, such as "left luggage" and "suspicious handover". A transfer learning optimization strategy is adopted, and the pre-trained model is fine-tuned on the Kinetics dataset. The loss function is a cross-entropy loss and an angular feature regularization term.

[0120] Examples of abnormal behavior identification include analysis of simulation scenarios, data processing procedures, parameter quantities and training duration, and analysis of behavior identification results in simulated port scenarios;

[0121] The simulation scenario was set up to simulate the port inspection area where a suspicious person was found squatting and touching the suitcase several times.

[0122] The data processing flow includes the following steps:

[0123] (1) The drone collects video and extracts skeletal data, which is then interpolated to 60 fps using the LI module;

[0124] (2) The HFBA module calculates the hand-elbow-shoulder angle features and detects abnormal squatting angles. The threshold for abnormal angles is set to >120°.

[0125] (3) SGA-GCN identifies the behavior as “suspicious item placement” and sets the abnormal behavior probability threshold to 85%;

[0126] (4) The alarm module is combined with regional authority. The regional authority is that the person is prohibited from staying in the inspection area, which triggers a level 1 alarm and pushes it to the law enforcement officer's terminal.

[0127] Parameter Quantity and Training Time Analysis: Table 1 shows a comparison of the parameter quantity and training time of the SGA-GCN model used in the present invention and the existing AAGCN model. As can be seen from Table 1, the SGA-GCN model used in the present invention reduces the parameter quantity and training time by more than 30%, indicating that the spatial grouping attention mechanism infers the attention map and reduces the network complexity, effectively reducing the training time of the model.

[0128] Table 1 Comparative analysis of the number of parameters and training time of the SGA-GCN model and the AAGCN model

[0129]

[0130] Analysis of behavior recognition results in a simulated port scenario: The SGA-GCN model incorporates different data branch combinations into the behavior recognition process, such as FJBMA, FJ joint features + FB skeletal features + FM spatiotemporal features + FA angular features. The analysis results are shown in Table 2. The high-order feature encoding of skeletal angles can capture subtle movement differences between joints, such as finger bending angles and leg swing amplitudes, significantly enhancing the model's discriminative ability. Angular features provide supplementary information beyond traditional joint and skeletal features, which is particularly important for distinguishing similar behaviors, such as "normal walking" and "walking with contraband."

[0131] Table 2 Analysis results of SGA-GCN model with different data branch combinations

[0132]

[0133] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A big data identification system for investigative behavior at smart ports, characterized by: include: S1: Data acquisition and transmission module, used to collect aerial video in real time through a drone cluster and transmit it to edge computing nodes through multi-node load balancing; S2: Linear interpolation module, used to dynamically insert intermediate frames into low-frame-rate videos to improve temporal information density; S3: High-order feature encoding module, used to extract the coordinates of human joints, calculate static angle features and dynamic angle change rate features, and generate a fusion feature matrix; S4: A lightweight graph convolutional network module, which is used to dynamically divide and group the motion amplitudes of the joints and enhance the key area features by combining the spatiotemporal attention mechanism; S5: Behavior classification and alarm module, used to trigger alarms based on behavior probability and record abnormal information.

2. The big data identification system for investigative behavior at smart ports according to claim 1 is characterized in that: The static angle features extracted by the high-order feature encoding module include adjacent node angles, central orientation angles and symmetrical node angles. The dynamic angle features are the rate of change of angles between adjacent frames, and noise is eliminated through Z-Score standardization and sliding average filtering.

3. The big data identification system for investigative behavior at smart ports according to claim 1 is characterized in that: The SGA-GCN module predicts joint importance scores through a lightweight fully connected layer, dynamically merges the joint points to generate groups based on Euclidean distance, and enhances local action features through spatial average feature weighting.

4. The method for applying to a big data identification system for investigative behavior at a smart port according to any one of claims 1 and 3, characterized in that: The steps include: Aerial video is collected by drones, and the linear interpolation algorithm is used to unify the number of frames to the target length; Extracting the coordinates of the human body joint points in each frame, and calculating the static angle feature and the dynamic angle change rate feature; Dynamically dividing the groups based on the motion amplitude of the joint points, and enhancing the key area features through the spatial grouping attention mechanism; The spatiotemporal features are integrated to classify behaviors, and an alarm is sent if the probability of abnormal behavior exceeds the threshold.

5. The method for a big data identification system for investigative behavior in a smart port according to claim 4 is characterized in that: The linear interpolation algorithm includes the following steps: S201: Calculate the number of frames that need to be inserted between adjacent frames: Number of frames to be deleted after inserting: f d =(f f +1)×f0-f u ; Among them, f u is the target frame number; f0 is the original frame number; ρ(x) is the rounding function; S202: The joint point detection algorithm extracts N joint point coordinates in each frame; S203: Calculate the displacement matrix between adjacent frames: M i =C i+1 -C i Among them, C i is the skeleton data extracted from the i-th frame of the original video; S204: Calculate the single-step displacement matrix between adjacent frames: P i =M i / f f ; S205: Generate the k-th vertex λ of the i-th frame i Coordinate vector of: C i,k =C i +(k-1)×P i ; Video data after interpolation F∈R C×T×N , where C is the coordinate dimension, T is the target frame length, i.e. 60, and N is the number of joint points.

6. The method for a big data identification system for investigative behavior in a smart port according to claim 4 is characterized in that: The dynamic angle change rate characteristic is expressed by the formula Calculate and combine with the static features to form a 6-dimensional fusion feature matrix; where R s,t (Ni) represents the static joint angle value of the i-th joint point and the t-th frame; Δt t Indicates the time interval between the t-th frame and the t-1-th frame.

7. The method of a big data identification system for investigative behavior in a smart port according to claim 4 is characterized in that: The dynamic grouping includes: normalizing the joint motion amplitude, generating the importance score through the Softmax function, screening the joint points whose scores exceed a threshold, and merging them into the dynamic grouping based on the Euclidean distance.

8. The method for a big data identification system for investigative behavior in a smart port according to claim 4 is characterized in that: The spatial grouping attention mechanism includes: calculating the spatial average features within the group, generating attention weights through the fully connected layer, weightedly enhancing the local action features and suppressing noise.

9. The method for a big data identification system for investigative behavior in a smart port according to claim 4 is characterized in that: The behavior classification adopts the joint optimization of cross entropy loss and angle feature regularization term, outputs the abnormal behavior probability distribution and sets 80% as the alarm threshold.

10. The method for a big data identification system for investigative behavior in a smart port according to claim 4 is characterized in that: The warning information includes the behavior type, timestamp and location coordinates, and is pushed to the law enforcement officer's terminal via SMS or platform interface.

Images