Track detection method and device based on edge perception and risk prediction, and medium

Abstract

The embodiment of the invention discloses a trajectory detection method and device based on edge perception and risk prediction and a medium, belongs to the technical field of trajectory anomaly detection, and solves the problem that in an existing security detection technology, misjudgment or missed judgment is likely to occur in a scene with the high requirement for security accuracy. Comprising the following steps: acquiring target image data acquired by a plurality of monitoring and photographing devices through a plurality of edge devices, and performing access feature vector extraction on the target image data; wherein the access feature vector at least comprises a visual feature vector and an access tag feature vector; based on the access feature vector, matching the multiple pieces of target image data to reconstruct a global access track corresponding to the target object; inputting the global access track and the access information into a multi-modal large model, and outputting a behavior risk score corresponding to the target object through the multi-modal large model; and based on the behavior risk score and the access feature vector, determining a trajectory risk level and alarm data, and pushing the alarm data to a security management system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of trajectory anomaly detection, in particular to a trajectory detection method and device based on edge perception and risk prediction and a medium. BACKGROUND

[0002] With the continuous upgrading of smart city construction and public safety demand, accurate monitoring and early warning of target object behavior risk have become the core challenge in the security field.

[0003] Traditional security systems have many limitations when dealing with complex scenarios. On the one hand, most systems rely on a single monitoring device to collect data, which is limited by fixed viewing angle and limited coverage, making it difficult to capture the complete cross-region activity trajectory of the target object, especially in large public places or multi-story buildings. Blind spots are likely to occur, leading to fragmented behavior analysis and making it difficult to form coherent behavior analysis basis.

[0004] On the other hand, existing risk assessment techniques are mostly based on single-modal data analysis, and a single analysis mode cannot fully exploit the complex characteristics of the target trajectory. In scenarios where security accuracy is required, false positives or false negatives are likely to occur. SUMMARY

[0005] The embodiments of the present application provide a trajectory detection method, device and medium based on edge perception and risk prediction, to solve the technical problem that existing security detection techniques cannot fully exploit the complex characteristics of the target trajectory, and are likely to produce false positives or false negatives in scenarios where security accuracy is required.

[0006] The embodiments of the present application adopt the following technical solutions: The embodiments of the present application provide a trajectory detection method based on edge perception and risk prediction. It includes obtaining target image data collected by multiple monitoring devices through multiple edge devices, and extracting access feature vectors from the target image data; wherein the access feature vectors include at least visual feature vectors and access label feature vectors; based on the access feature vectors, matching multiple target image data to reconstruct the global access trajectory corresponding to the target object; inputting the global access trajectory and access information into a multi-modal large model to output the behavior risk score corresponding to the target object through the multi-modal large model; based on the behavior risk score and the access feature vectors, determining the trajectory risk level and the alarm data, and pushing the alarm data to the security management system.

[0007] The embodiments of the present application realize accurate matching of multi-source image data by multiple edge devices cooperatively obtaining different target image data and extracting visual and access label feature vectors, complete restoration of the global access track of a target object across regions and for a long time, and elimination of monitoring blind spots. Secondly, deep analysis is performed through a multi-modal large model to avoid one-sidedness of single-modal data. The identification accuracy of abnormal behavior and disguised behavior is greatly improved through model mining of potential correlation features of target behavior, reducing misjudgment and omission, meeting the needs of high-security scenarios. In addition, the embodiments of the present application output behavior risk scores based on global track and multi-modal features, and dynamically determine risk levels in combination with access feature vectors, quickly respond to changes in target behavior, and timely generate alarm data and push it to a security management system, thereby reducing security risks and improving the timeliness and accuracy of security decisions.

[0008] In an implementation manner of the present application, a plurality of edge devices obtain target image data collected by a plurality of monitoring devices, and access feature vectors are extracted from the target image data, specifically including: receiving, by the plurality of edge devices, target image data collected by the plurality of monitoring devices respectively; performing target detection on the target image data by a YOLO model corresponding to the plurality of edge devices to obtain a recognition box corresponding to a target object; when it is determined that the target object in the recognition box does not belong to a preset reference list table, performing visual feature vector extraction on the target image data by a ReID model; and generating an access label feature vector based on regional label information corresponding to the target image data.

[0009] In an implementation manner of the present application, based on the access feature vectors, the plurality of target image data are matched to reconstruct a global access track corresponding to a target object, specifically including: obtaining access feature vectors uploaded by a plurality of edge devices respectively; comparing the access feature vectors uploaded by different edge devices respectively to determine a similarity; based on the similarity, constructing a similar image data set corresponding to the same target object; sorting the similar image data set based on time stamps corresponding to each target image data in the similar image data set; based on the arrangement order and the access label feature vectors corresponding to each target image data, constructing an access path graph corresponding to the target object by a graph neural network to obtain the global access track based on the access path graph.

[0010] In an implementation manner of the present application, before the global access trajectory and the access information are input into the multi-modal large model, the method further comprises: in the case of network interruption or missing trajectory, reconstructing the access trajectory sequence by the Mamba network and the Transformer-based time sequence model; constructing a graph neural network by taking each monitoring device as a node and determining an edge weight based on the historical trajectory transfer frequency; performing path prediction by the graph neural network and the reconstructed access trajectory sequence; determining a node transfer probability based on the graph attention mechanism, and generating a plurality of reference paths based on the node transfer probability to complete the global access path based on the plurality of reference paths.

[0011] In an implementation manner of the present application, the access trajectory sequence is reconstructed by the Mamba network and the Transformer-based time sequence model, specifically comprising: constructing a sliding time window and marking the missing frames corresponding to the target image with a mask to form a masked input sequence; extracting trajectory dynamic features by the selective state space model of the Mamba network, and weighting the contribution degrees of different time steps by the attention mechanism to obtain first prediction features corresponding to the missing frames; determining the time sequence correlation degree by the position encoding and multi-head self-attention mechanism of the Transformer-based time sequence model, and generating second prediction features corresponding to the missing frames based on the time sequence correlation degree; and weighting and fusing the first prediction features and the second prediction features to reconstruct the access trajectory sequence.

[0012] In an implementation manner of the present application, the global access path is completed based on the plurality of reference paths, specifically comprising: determining the prior probability corresponding to each reference path based on the historical trajectory data; determining the confidence of the reference path based on the prior probability; when the confidence of the reference path meets a pre-set confidence threshold, completing the global access path based on the reference path; otherwise, generating a path question warning according to the confidence.

[0013] In an implementation manner of the present application, the global access trajectory and the access information are input into the multi-modal large model to output the behavior risk score corresponding to the target object by the multi-modal large model, specifically comprising: inputting the global access trajectory and the access information into the multi-modal large model; wherein the access information at least includes access time window and region permission information; performing semantic analysis on the global access trajectory and the access information by the multi-modal large model and the pre-set prompt word to obtain semantic information corresponding to the behavior intention of the target object; determining the behavior semantic label matched with the semantic information in the pre-set label library; determining the risk score corresponding to the behavior intention of the target object based on the semantic label and the corresponding risk score table.

[0014] In an implementation form of the present application, based on the behavior risk score and the access feature vector, the trajectory risk level and the alarm data are determined, specifically comprising: based on the access feature vector, determining the stay time and the access frequency of the target object corresponding to different regions respectively; and determining the access permission level corresponding to different regions respectively; based on the stay time, the access frequency, the access permission level and the risk score, triggering different early warning levels and sending the alarm data; wherein the alarm data at least includes the access frequency, the region category, the behavior label and the time feature.

[0015] The embodiment of the present application provides a trajectory detection device based on edge perception and risk prediction, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: acquire target image data collected by a plurality of monitoring devices through a plurality of edge devices, and extract access feature vectors from the target image data; wherein the access feature vectors at least include visual feature vectors and access label feature vectors; based on the access feature vectors, match the plurality of target image data to reconstruct a global access trajectory corresponding to a target object; input the global access trajectory and access information into a multi-modal large model to output a behavior risk score corresponding to the target object through the multi-modal large model; based on the behavior risk score and the access feature vectors, determine a trajectory risk level and alarm data, and push the alarm data to a security management system.

[0016] The embodiment of the present application provides a non-volatile computer storage medium, which stores computer executable instructions, and the computer executable instructions are configured to: acquire target image data collected by a plurality of monitoring devices through a plurality of edge devices, and extract access feature vectors from the target image data; wherein the access feature vectors at least include visual feature vectors and access label feature vectors; based on the access feature vectors, match the plurality of target image data to reconstruct a global access trajectory corresponding to a target object; input the global access trajectory and access information into a multi-modal large model to output a behavior risk score corresponding to the target object through the multi-modal large model; based on the behavior risk score and the access feature vectors, determine a trajectory risk level and alarm data, and push the alarm data to a security management system.

[0017] The at least one technical solution adopted by the embodiments of the present application can achieve the following beneficial effects: the embodiments of the present application realize accurate matching of multi-source image data, complete restoration of the global access track of a target object across regions and for a long time, and elimination of monitoring blind spots by cooperatively obtaining different target image data through multiple edge devices and extracting visual and access label feature vectors. Secondly, deep analysis is performed through a multi-modal large model to avoid one-sidedness of single-modal data. The identification accuracy of abnormal behavior and disguised behavior is greatly improved through model mining of potential correlation features of target behavior, reducing misjudgment and omission, meeting the needs of high-security scenarios. In addition, the embodiments of the present application output behavior risk scores based on global tracks and multi-modal features, and dynamically determine risk levels in combination with access feature vectors, quickly respond to changes in target behavior, and timely generate alarm data and push it to a security management system, thereby reducing security risks and improving the timeliness and accuracy of security decisions. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can be obtained by those skilled in the art without creative labor. In the drawings: Figure 1 A flowchart of a trajectory detection method based on edge perception and risk prediction is provided for the embodiments of the present application. Figure 2 A structural schematic diagram of a trajectory detection device based on edge perception and risk prediction is provided for the embodiments of the present application.

[0019] Reference signs: 200: trajectory detection device based on edge perception and risk prediction, 201: processor, 202: memory. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a trajectory detection method, device and medium based on edge perception and risk prediction.

[0021] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0022] The trajectory detection system based on edge perception and risk prediction corresponding to the embodiment of the application is composed of five layers of architecture: The perception layer: camera network, face recognition module, license plate recognition module; The edge computing layer: YOLO model, feature extraction model, preliminary trajectory modeling; The intermediate processing layer: data fusion, image embedding, feature compression; The intelligent analysis layer: multi-modal large model, trajectory reconstruction, behavior analysis; The early warning and interface layer: risk prompt, alarm display, API interface.

[0023] The technical solutions of the embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0024] Figure 1 A trajectory detection method based on edge perception and risk prediction provided by the embodiment of the application is shown in the flow chart as Figure 1 The trajectory detection method based on edge perception and risk prediction includes the following steps: S101, obtaining target image data collected by a plurality of monitoring devices through a plurality of edge devices, and extracting access feature vectors from the target image data.

[0025] In an implementation manner of the application, the target image data collected by a plurality of monitoring devices is received through a plurality of edge devices. The target image data is detected by the corresponding YOLO model of the plurality of edge devices to obtain the recognition box corresponding to the target object. In the case where the target object in the recognition box does not belong to the preset reference list, the visual feature vector of the target image data is extracted by the ReID model. And, the access label feature vector is generated based on the region label information corresponding to the target image data. The access feature vector at least includes the visual feature vector and the access label feature vector.

[0026] Specifically, a plurality of monitoring devices are distributed in different areas and continuously collect image data containing target objects. These monitoring devices transmit the collected target image data to the corresponding plurality of edge devices in real time. Each edge device is equipped with a YOLO (You Only Look Once) model, and the edge device inputs the received target image data into the YOLO model. The model will detect each type of object in the image quickly and generate a corresponding recognition box for each detected target object.

[0027] Further, after obtaining the bounding box, it is judged whether the target object in the bounding box belongs to a preset reference list. If the target object is not in the preset reference list, it is an unknown object that needs to be focused on. At this time, the ReID model is called to process the target image data. It should be noted that the ReID model is used to extract the visual features of the target object, and by analyzing the visual information such as appearance, texture, shape, etc. of the target object, a corresponding visual feature vector is generated.

[0028] Further, according to the region label information corresponding to the target image data, an access label feature vector is generated. The region label information contains specific location, entry time, region attribute, such as important region, restricted region and other information of the target object. These information is encoded and integrated into an access label feature vector.

[0029] S102, based on the access feature vector, the plurality of target image data is matched to reconstruct the global access trajectory corresponding to the target object.

[0030] In an implementation manner of the present application, the access feature vectors uploaded by a plurality of edge devices are obtained. The access feature vectors uploaded by different edge devices are compared to determine the similarity. Based on the similarity, a similar image data set corresponding to the same target object is constructed. Based on the time stamps corresponding to each target image data in the similar image data set, the similar image data set is sorted. Based on the arrangement order and the access label feature vectors corresponding to each target image data, an access path graph corresponding to the target object is constructed through a graph neural network to obtain a global access trajectory based on the access path graph.

[0031] Specifically, the plurality of edge devices upload the access feature vectors processed by themselves, including visual features and access label features, to the central processing system. The central processing system compares the access feature vectors uploaded by different edge devices, and quantifies the similarity between features through cosine similarity. The image targets corresponding to the features with a similarity greater than a preset similarity threshold are determined as the same target object. Based on the similarity calculation result, all target image data belonging to the same target object are collected into a similar image data set. The data set integrates the multi-dimensional features of the target object in different monitoring areas and at different time points.

[0032] Further, the similar image data set is arranged in time sequence by using the time stamp information of each target image data in the data set. The time stamp records the entry and exit time of the target object in each monitoring area, and through the sorting, the order of the target object movement can be restored, avoiding the disorder of the trajectory time sequence.

[0033] Further, based on the time-ordered image data, combined with the access label feature vector, an access path graph is constructed through a graph neural network, i.e., each spatio-temporal node is taken as a node of the graph, and the node attribute contains the visual feature and the access label. The movement relationship of the target object between adjacent nodes is taken as an edge of the graph, and the weight of the edge is determined by the time interval and the spatial distance. Finally, the global access trajectory of the target object is generated through graph structure reasoning, and the complete movement path and behavior pattern of the target object in the monitoring range are presented.

[0034] In an implementation manner of the present application, in the case of network interruption or trajectory missing, the access trajectory sequence is reconstructed through the Mamba network and the Transformer-based time sequence model. The monitoring device is taken as a node, a graph neural network is constructed, and the edge weight is determined based on the historical trajectory transfer frequency. The path is predicted through the graph neural network and the reconstructed access trajectory sequence. The node transfer probability is determined based on the graph attention mechanism, and the multiple reference paths are generated based on the node transfer probability, so as to complete the global access path based on the multiple reference paths.

[0035] Specifically, when the network interruption causes the data transmission failure or the monitoring device misses the trajectory, the system starts the Mamba network and the Transformer-based time sequence model to reconstruct the trajectory.

[0036] It should be noted that the Mamba network is a high-efficiency long-sequence modeling tool, which can capture the time sequence dynamic pattern in the historical trajectory; the Transformer model can mine the long-distance dependency relationship between the trajectory points through the self-attention mechanism. Through the combination of the two, the access trajectory sequence in the missing period is predicted and filled based on the existing complete trajectory segment, and the potential movement path of the target object is restored.

[0037] Further, the monitoring device is taken as a node of the graph neural network, and a topological structure is constructed. By analyzing the transfer frequency of the target object between different monitoring devices, the edge weight between nodes is calculated, i.e., the higher the transfer frequency, the greater the edge weight, and the higher the probability of the movement of the target object between the corresponding monitoring devices. The reconstructed access trajectory sequence is input into the graph neural network, and the path is deduced combined with the edge weight between nodes. The graph neural network propagates the trajectory feature between nodes through the message passing mechanism, and learns the movement pattern of the target object. The model predicts the monitoring device node that may be reached at the next moment based on the end node state of the current reconstructed trajectory, combined with the historical transfer law, and generates the potential movement path in the future period.

[0038] Further, the transition probability of each node is calculated by using a graph attention mechanism, and a plurality of possible reference paths are generated based on the probability distribution, each path corresponding to a potential movement trend. The missing sections in the global access trajectory are completed according to the reference path, ensuring the continuity and integrity of the trajectory.

[0039] In an implementation manner of the present application, the reconstruction process of the access trajectory sequence specifically includes: constructing a sliding time window and marking the missing frames corresponding to the target image with a mask to form a masked input sequence. The trajectory dynamic features are extracted by a selective state space model of the Mamba network, and the contribution degrees of different time steps are weighted by an attention mechanism to obtain the first prediction features corresponding to the missing frames. The time sequence correlation degrees are determined by the position encoding and the multi-head self-attention mechanism of the Transformer-based time sequence model, and the second prediction features corresponding to the missing frames are generated based on the time sequence correlation degrees. The first prediction features and the second prediction features are weighted and fused to reconstruct the access trajectory sequence.

[0040] Specifically, the sliding window of the fixed time span is first fixed in the embodiment of the present application, which is used to intercept the time sequence of the target image. For the missing frames in the sequence, the positions thereof are marked in the form of a mask to form a masked input sequence. The selective state space model of the Mamba network is used to process the masked input sequence. The selective state space model models the physical motion law of the trajectory by a linear dynamic system, and can capture the dynamic change pattern of the trajectory in the time dimension. At the same time, the contribution degrees of different time steps are weighted by an attention mechanism, and higher weight is given to the recent trajectory points, so as to generate the first prediction features corresponding to the missing frames.

[0041] Further, the input sequence is processed twice by the Transformer-based time sequence model: first, the absolute time sequence information is embedded for each time step by using the position encoding, and then the correlation weights between the trajectory points at different time steps are calculated by the multi-head self-attention mechanism to capture the long-distance time sequence dependence. The second prediction features of the missing frames are generated based on the correlation degrees by the model.

[0042] Further, the dynamic features output by the Mamba network and the time sequence correlation features output by the Transformer are weighted and fused. The weight parameters are determined according to the historical data, for example, the Mamba feature weight is increased in the straight line motion scene, and the Transformer feature weight is increased in the complex turning scene. The fused features are decoded to complete the missing frames, and the complete reconstruction of the access trajectory sequence is realized.

[0043] In an implementation form of the present application, the global access path is completed based on a plurality of reference paths, specifically comprising: determining the prior probability corresponding to each reference path based on historical trajectory data. The confidence of the reference path is determined based on the prior probability. When the confidence of the reference path meets a preset confidence threshold, the global access path is completed based on the reference path. Otherwise, a path question warning is generated according to the confidence.

[0044] Specifically, the moving data of the target object in a similar scene is extracted from the historical trajectory database, and the frequency of each reference path in the historical data is counted. For example, if a certain reference path appears 30 times in 100 historical trajectories, its prior probability is 0.3. After obtaining the prior probability, the environmental factors of the current scene, such as real-time crowd density, regional opening state, etc., are further integrated to correct the probability to obtain the final confidence. For example, if the prior probability of a certain reference path is 0.3, but the current monitoring area is temporarily closed due to construction, its confidence will be downgraded to 0.15 through an environmental weight factor. If the confidence of a certain reference path reaches or exceeds the threshold, the system determines that the path is a high-reliability path, and its nodes are embedded in the missing section of the global access trajectory in chronological order to complete the trajectory completion. If the confidence of all reference paths is lower than the threshold, it means that there is a large uncertainty in the possible path of the current missing trajectory, which may be caused by abnormal behavior or data missing. At this time, a path question warning is generated, the missing section is marked, and the potential risk is prompted.

[0045] S103, inputting the global access trajectory and the access information into a multi-modal large model to output a behavior risk score corresponding to the target object through the multi-modal large model.

[0046] In an implementation form of the present application, the global access trajectory and the access information are input into a multi-modal large model; wherein the access information at least includes access time window and regional permission information. The semantic information corresponding to the behavior intention of the target object is obtained by performing semantic analysis on the global access trajectory and the access information through the multi-modal large model and the preset prompt word. In the preset label library, the behavior semantic label matching the semantic information is determined. Based on the semantic label and the corresponding risk score table, the risk score corresponding to the behavior intention of the target object is determined.

[0047] Specifically, the global access trajectory and the access information of the target object are input into the multi-modal large model. The access time window records the entry and exit time of the target object in each region, and the regional permission information identifies the access restriction of different regions. The multi-modal large model is guided to perform semantic analysis on the input data through the preset prompt word. The model combines the spatio-temporal characteristics of the global access trajectory and the permission logic of the access information to extract the semantic information corresponding to the behavior intention of the target object.

[0048] Further, the behavior semantic label matching the parsed semantic information is searched in a preset label library. The label library contains preset behavior pattern classification, and each label corresponds to specific semantic definition and feature description. The abstract semantic information is mapped to specific labels through semantic similarity calculation or keyword matching. According to the matched behavior semantic label, the corresponding risk score table is queried to calculate the risk score of the behavior intention of the target object.

[0049] In S104, the trajectory risk level and the alarm data are determined based on the behavior risk score and the access feature vector, and the alarm data is pushed to the security management system.

[0050] In an implementation manner of the present application, the access time and the access frequency of the target object in different regions are determined based on the access feature vector. The access permission level corresponding to each region is determined. Based on the access time, the access frequency, the access permission level and the risk score, different early warning levels are triggered, and the alarm data is sent. The alarm data at least includes the access frequency, the region category, the behavior label and the time feature.

[0051] Specifically, the access time of the target in a specific region is calculated through the visual feature and the timestamp information of the access label. The access frequency of the target to each region within a certain period is counted through cross-device trajectory matching. The access permission level of different monitoring regions is allocated according to the region attribute preset by the security system. The access time, the access frequency, the permission level and the behavior risk score are used as input factors to trigger different early warning levels through a preset risk assessment model.

[0052] For example, if the target stays in the normal region for a normal time, but the access frequency is 2 times the historical average, a low-level early warning is triggered. If the target enters the forbidden area without authorization and stays for more than 5 minutes, combined with a high risk score, a highest-level early warning is triggered directly. The alarm data is sent to the security management system to assist the management personnel to quickly locate the risk point and take measures.

[0053] Figure 2 A structural schematic diagram of a trajectory detection device based on edge perception and risk prediction is provided for the embodiments of the present application. As shown in Figure 2As shown, the trajectory detection device 200 based on edge awareness and risk prediction comprises: at least one processor 201; and a memory 202 in communication connection with the at least one processor 201; wherein the memory 202 stores instructions executable by the at least one processor 201, and the instructions are executed by the at least one processor 201 to enable the at least one processor 201 to: acquire target image data collected by a plurality of monitoring devices through a plurality of edge devices, and extract access feature vectors from the target image data; wherein the access feature vectors at least include visual feature vectors and access label feature vectors; match a plurality of target image data based on the access feature vectors to reconstruct a global access trajectory corresponding to a target object; input the global access trajectory and access information into a multi-modal large model to output a behavior risk score corresponding to the target object through the multi-modal large model; and determine a trajectory risk level and alarm data based on the behavior risk score and the access feature vectors, and push the alarm data to a security management system.

[0054] The non-volatile computer storage medium provided by the embodiments of the present application stores computer executable instructions, and the computer executable instructions are configured to: acquire target image data collected by a plurality of monitoring devices through a plurality of edge devices, and extract access feature vectors from the target image data; wherein the access feature vectors at least include visual feature vectors and access label feature vectors; match a plurality of target image data based on the access feature vectors to reconstruct a global access trajectory corresponding to a target object; input the global access trajectory and access information into a multi-modal large model to output a behavior risk score corresponding to the target object through the multi-modal large model; and determine a trajectory risk level and alarm data based on the behavior risk score and the access feature vectors, and push the alarm data to a security management system.

[0055] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other. Each embodiment mainly describes the difference from other embodiments. Especially, the device, equipment and non-volatile computer storage medium embodiments are basically similar to the method embodiments, so the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0056] The above only describes the embodiments of the present application and is not used to limit the present application. The embodiments of the present application can be variously changed and modified by those skilled in the art. The modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A trajectory detection method based on edge perception and risk prediction, characterized in that: The method comprises: Acquire target image data collected by multiple monitoring devices through multiple edge devices, and extract access feature vectors from the target image data; wherein the access feature vectors include at least visual feature vectors and access label feature vectors; Matching the plurality of target image data based on the access feature vector to reconstruct a global access trajectory corresponding to the target object; Inputting the global access trajectory and access information into a multimodal large model to output a behavior risk score corresponding to the target object through the multimodal large model; Based on the behavior risk score and the access feature vector, a trajectory risk level and alarm data are determined, and the alarm data is pushed to a security management system.

2. The trajectory detection method based on edge perception and risk prediction according to claim 1, characterized in that: The method of acquiring target image data collected by multiple monitoring devices through multiple edge devices and extracting access feature vectors of the target image data specifically includes: Receiving target image data respectively collected by the plurality of monitoring devices through the plurality of edge devices; Performing target detection on the target image data using the YOLO models corresponding to the multiple edge devices to obtain an identification box corresponding to the target object; When it is determined that the target object in the recognition frame does not belong to the preset reference list, extracting a visual feature vector from the target image data using a ReID model; And, based on the region label information corresponding to the target image data, the access label feature vector is generated.

3. The trajectory detection method based on edge perception and risk prediction according to claim 1, characterized in that: The matching of the plurality of target image data based on the access feature vector to reconstruct a global access trajectory corresponding to the target object specifically includes: Obtaining access feature vectors uploaded by a plurality of edge devices respectively; Comparing the access feature vectors uploaded by different edge devices to determine similarity; Based on the similarity, constructing a similar image dataset corresponding to the same target object; sorting the similar image dataset based on timestamps corresponding to the target image data in the similar image dataset; Based on the arrangement order and the access label feature vector corresponding to each target image data, an access path graph corresponding to the target object is constructed through a graph neural network to obtain the global access trajectory based on the access path graph.

4. The trajectory detection method based on edge perception and risk prediction according to claim 1, characterized in that: Before inputting the global access trajectory and access information into the multimodal large model, the method further includes: In the event of network interruption or trajectory loss, the access trajectory sequence is reconstructed through the Mamba network and Transformer-based timing model; Each of the surveillance devices is used as a node to construct a graph neural network, and the outgoing edge weight is determined based on the historical trajectory transfer frequency; Performing path prediction using the graph neural network and the reconstructed access trajectory sequence; A node transition probability is determined based on a graph attention mechanism, and multiple reference paths are generated based on the node transition probability, so as to complete the global access path based on the multiple reference paths.

5. The trajectory detection method based on edge perception and risk prediction according to claim 4 is characterized in that: The access trajectory sequence is reconstructed using the Mamba network and the Transformer-based timing model, specifically including: Constructing a sliding time window and masking the missing frames corresponding to the target image to form a masked input sequence; Extracting trajectory dynamic features through the selective state space model of the Mamba network, and weighting the contributions of different time steps through the attention mechanism to obtain the first prediction feature corresponding to the missing frame; Determining a temporal correlation by using the position encoding and multi-head self-attention mechanism of the Transformer-based temporal model, and generating a second prediction feature corresponding to the missing frame based on the temporal correlation; The first prediction feature and the second prediction feature are weightedly fused to reconstruct the access trajectory sequence.

6. The trajectory detection method based on edge perception and risk prediction according to claim 4, characterized in that: The constructing and completing the global access path based on the multiple reference paths specifically includes: Determining the prior probability corresponding to each of the reference paths based on the historical trajectory data; Determining the confidence level of the reference path based on the prior probability; When the confidence of the reference path meets a preset confidence threshold, completing the global access path based on the reference path; Otherwise, a path question warning is generated according to the confidence level.

7. The trajectory detection method based on edge perception and risk prediction according to claim 1, characterized in that: Inputting the global access trajectory and access information into a multimodal large model to output a behavior risk score corresponding to the target object through the multimodal large model specifically includes: Inputting the global access trajectory and access information into a multimodal macro model; wherein the access information includes at least access time window and regional authority information; Performing semantic analysis on the global access trajectory and the access information using the multimodal large model and preset prompt words to obtain semantic information corresponding to the target object's behavioral intention; Determining a behavior semantic tag that matches the semantic information in a preset tag library; Based on the semantic tags and the corresponding risk score table, a risk score corresponding to the behavioral intention of the target object is determined.

8. The trajectory detection method based on edge perception and risk prediction according to claim 1, characterized in that: Determining the trajectory risk level and warning data based on the behavior risk score and the access feature vector specifically includes: Based on the visit feature vector, determining the residence time and visit times of the target object in different areas respectively; and determining the access rights levels corresponding to the different areas; Based on the residence time, the number of visits, the access permission level, and the risk score, trigger different warning levels and send warning data; The alarm data includes at least the number of visits, area categories, behavior tags and time characteristics.

9. Trajectory detection equipment based on edge perception and risk prediction, characterized in that: The device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute the method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 8.

Images