Track file construction method and system

Abstract

The invention relates to the technical field of trajectory data analysis and processing, and discloses a trajectory file construction method and system.The trajectory file construction method comprises the steps of multi-dimensional trajectory feature decomposition and extraction, specifically, suspect trajectory elements are divided into the time dimension, the space dimension and the behavior dimension through multi-level feature decomposition; a trajectory element time sequence dependency graph construction step: constructing a directed acyclic graph to represent a time sequence relationship between elements; a step of constructing a dynamic system model of an ordinary differential equation, wherein trajectory evolution dynamics is expressed by using an ordinary differential equation parameterized by a neural network; and a continuous and discrete mixed archive structure construction step: combining discrete observation data with a continuous dynamic model, and constructing an archive structure with efficient compression storage and multi-level indexing. According to the method, the continuous evolution process of the suspect trajectory can be accurately expressed, trajectory state inference of any time granularity is supported, and the accuracy of trajectory analysis and prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of trajectory data analysis and processing, and more specifically, to a trajectory archive construction method and system. Background Art

[0002] With the rapid development of technologies such as intelligent monitoring, electronic fences, and mobile positioning, the collection of suspect trajectory data has become more convenient and comprehensive, and a large amount of trajectory data has provided important clues for criminal case investigation. However, in the face of the continuous growth of trajectory data, the existing trajectory archive construction methods have the following technical problems:

[0003] Traditional methods treat trajectory data as a whole, lacking multi-dimensional feature decomposition and semantic understanding, resulting in difficulty in distinguishing key features from non-key features, low retrieval efficiency and redundant storage; existing technologies ignore the temporal dependencies between trajectory elements and only focus on single trajectory points or simple sequences, resulting in the inability to efficiently reconstruct complex trajectory behavior patterns during archive retrieval and analysis, especially in the analysis of suspect activities over a long time span; most trajectory analysis methods use discrete temporal dependency graphs or simple statistical models, which cannot accurately express the trajectory evolution dynamics under continuous time, resulting in insufficient accuracy when analyzing rapidly changing dynamic scenes, especially in predicting the potential trajectory of suspects.

[0004] Traditional trajectory archive storage structures usually process data by overall compression or simple sampling, which cannot achieve refined management at the element level and is difficult to meet the dual needs of efficient retrieval and accurate expression at the same time, especially in long-term monitoring and complex case analysis scenarios. Although some studies have tried to use deep learning methods to improve trajectory data processing, most of these methods focus on trajectory prediction or simple classification in specific scenarios, lack a systematic trajectory archive construction framework, and are difficult to support complex trajectory analysis and query needs. At the same time, these methods usually require a large amount of labeled data for training, which is difficult to adapt to the characteristics of sparse data and changing scenarios in public security operations.

[0005] Therefore, there is an urgent need for a suspect trajectory archive construction method that can decompose trajectory features in multiple dimensions, capture the temporal dependencies between elements, accurately model continuous-time trajectory dynamics, and construct an efficient index and compression storage structure to support efficient and accurate suspect trajectory analysis and prediction. Summary of the invention

[0006] The present invention provides a trajectory archive construction method and system, which solve the technical problems in the related art of incomplete trajectory feature extraction, insufficient expression of temporal dependency, inaccurate continuous dynamics modeling and low efficiency of archive structure retrieval.

[0007] The present invention provides a trajectory archive construction method, comprising the following steps:

[0008] Obtain the original trajectory data, perform multi-dimensional feature decomposition on the original trajectory data, and generate a time dimension feature set, a space dimension feature set, and a behavior dimension feature set;

[0009] Based on the feature set, a temporal dependency graph of trajectory elements is constructed. The temporal dependency graph of trajectory elements is represented as a directed acyclic graph to capture the temporal relationship between trajectory elements.

[0010] Using the constructed trajectory element time-series dependency graph, a trajectory dynamic system model based on Neural ODEs is constructed, expressed as a continuous-time dynamic system, to achieve trajectory state evolution modeling; the Neural ODE trajectory dynamic system model is used to achieve trajectory state inference at any time point, and supports the recovery of continuous trajectories from finite observation points;

[0011] Construct a continuous-discrete hybrid archive structure to simultaneously save discrete observation points and continuous dynamic models, apply an adaptive compression storage strategy, and build a multi-level indexing system.

[0012] In a preferred embodiment, in the multi-dimensional feature decomposition step, time dimension feature extraction includes:

[0013] Apply the time segmentation algorithm to extract time series features, including frequency features, duration features, and periodic features, to form a time feature vector , which is calculated as:

[0014] ;

[0015] in is the time feature extraction function, represents the original trajectory data, For the time characteristics, Represents the total number of temporal features.

[0016] In a preferred embodiment, in the multidimensional feature decomposition step, the spatial dimension feature extraction includes: using a spatial trajectory analysis algorithm to extract spatial pattern features from a trajectory point sequence, including stop points, moving paths and activity areas, to form a spatial feature vector , which is calculated as:

[0017] ;

[0018] in is the spatial feature extraction function, represents the original trajectory data, For the spatial features, Represents the total number of spatial features.

[0019] In a preferred embodiment, in the multi-dimensional feature decomposition step, the behavior dimension feature extraction includes: using a semantic tagging algorithm to add behavior tags to the trajectory points, identifying suspicious behavior patterns, and forming a behavior feature vector , which is calculated as:

[0020] ;

[0021] in is the behavior feature extraction function, represents the original trajectory data, For the behavioral characteristics, Indicates the total number of behavioral characteristics.

[0022] In a preferred embodiment, the step of constructing the trajectory element temporal dependency graph includes: combining the feature vector set into a trajectory element set:

[0023] ;

[0024] in, represents a set of trajectory features, , , Respectively represent the 1st, 2nd, Trajectory elements, represents the number of trajectory elements, where the trajectory elements are represented as time feature vectors, space feature vectors or behavior feature vectors;

[0025] Constructing a directed acyclic graph ;in is the trajectory element set, is the temporal dependency edge set between elements, is the edge weight matrix; apply the graph optimization algorithm to perform sparse processing, remove low-correlation edges, and generate an optimized timing dependency graph ;in is the trajectory element set, is the optimized temporal dependency edge set between elements, is the optimized edge weight matrix;

[0026] An incremental graph update algorithm is adopted. When new trajectory elements are added, the graph structure is updated through local adjustments.

[0027] In a preferred embodiment, the sparse processing of the graph optimization algorithm is based on the following formula:

[0028] ;

[0029] in Indicates that from the element To feature The directed edge of represents the set of temporal dependency edges between optimized elements, is the temporal dependency edge set between elements, represents the edge weight, is the threshold parameter.

[0030] In a preferred embodiment, the trajectory dynamic system model of the Neural Ordinary Differential Equation is expressed as:

[0031] ;

[0032] in for Dimension time The trajectory state vector, The parameters are The neural network, represents the rate of change of the trajectory state vector over time, is the dimension of the state vector.

[0033] In a preferred embodiment, the trajectory dynamic system model of the Neural Ordinary Differential Equation includes a hidden state inference algorithm, which realizes trajectory state inference at any time point by solving the differential equation through numerical integration:

[0034] ;

[0035] in Initial time The trajectory state vector, To infer time The trajectory state vector, The parameters are The neural network, Indicates from time arrive 's points.

[0036] In a preferred embodiment, the step of constructing a continuous-discrete hybrid archive structure comprises:

[0037] Combine the collection of trajectory elements, the timing dependency graph, and the dynamic system model into a unified archive structure:

[0038] ;

[0039] Apply adaptive compression storage algorithm to compress feature collection, focusing on retaining key nodes and behavioral features;

[0040] Build a multi-level index system, including time index, space index, behavior pattern index and time-dependent path index;

[0041] Build trajectory archive query interface and restoration algorithm to support restoration of complete trajectory information from stored compressed elements and dynamic system models;

[0042] in, Represents the file structure, is the trajectory element set, is the optimized timing dependency graph. The parameters are Dynamic system model.

[0043] In a preferred embodiment, a trajectory file construction system is used to execute the above-mentioned trajectory file construction method, including:

[0044] A multi-dimensional feature decomposition module is used to perform multi-dimensional feature decomposition on the original trajectory data to generate a time dimension feature set, a space dimension feature set, and a behavior dimension feature set;

[0045] The timing dependency graph construction module is used to construct the trajectory element timing dependency graph. The trajectory element timing dependency graph is represented as a directed acyclic graph to capture the timing relationship between trajectory elements.

[0046] The Neural Ordinary Differential Equation Model module is used to construct the trajectory dynamic system model, expressed as a continuous-time dynamic system, and realize the evolution modeling of trajectory states;

[0047] The hybrid archive structure module is used to construct a continuous-discrete hybrid archive structure, save discrete observation points and continuous dynamic models at the same time, apply adaptive compression storage strategies, and build a multi-level index system.

[0048] The beneficial effects of the present invention are: by capturing the temporal dependency between trajectory elements, the constructed archive can more completely preserve the dynamic characteristics of the suspect's behavior, providing richer information support for trajectory analysis and behavior prediction;

[0049] By introducing continuous dynamics modeling, we can accurately capture and express the continuous evolution of the suspect's trajectory, support trajectory analysis and simulation prediction at any time granularity, and improve prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of a trajectory archive construction method of the present invention. DETAILED DESCRIPTION

[0051] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the present specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0052] At least one embodiment of the present invention discloses a method for constructing a trajectory archive, such as Figure 1 As shown, the following steps are included:

[0053] Step 1: obtain the original trajectory data, perform multi-dimensional feature decomposition on the original trajectory data, and generate a time dimension feature set, a space dimension feature set, and a behavior dimension feature set;

[0054] This step uses a multi-level feature decomposition algorithm to analyze the original trajectory data and generate a hierarchical feature structure, which includes:

[0055] Step 1.1, obtain the original trajectory data (in Represents a track point containing timestamp, geographic location and behavior label. represents the total number of trajectory points) is input into the multidimensional feature decomposition model and outputs three feature sets: time dimension feature set , spatial dimension feature set and behavioral dimension feature sets ;

[0056] The implementation of the multi-dimensional feature decomposition model includes three parallel branch networks: temporal feature extraction branch, spatial feature extraction branch, and behavioral feature extraction branch. In the scenario where the suspect frequently enters and exits a specific area, the model can simultaneously capture temporal patterns (such as visiting every Friday night), spatial patterns (such as short stays in remote corners), and behavioral characteristics (such as short-term contact with multiple people);

[0057] Step 1.2: For the time dimension features, apply the time segmentation algorithm to extract the time series features, including frequency features, duration features and periodic features, to form a time feature vector , the temporal eigenvector is calculated as follows:

[0058]

[0059] in is the time feature extraction function, represents the original trajectory data, For the time characteristics, Represents the total number of temporal features.

[0060] The time segmentation algorithm uses an adaptive window mechanism to analyze activities at different time scales. For example, for a suspect's daily and nighttime activity patterns, the algorithm can identify abnormal activity patterns during specific time periods (such as 2-4 a.m.).

[0061] Step 1.3: For the spatial dimension features, a spatial trajectory analysis algorithm is used to extract spatial pattern features from the trajectory point sequence, including the stop points, movement paths, and activity areas, to form a spatial feature vector , the spatial eigenvector is calculated as follows:

[0062]

[0063] in is the spatial feature extraction function, represents the original trajectory data, For the spatial features, Represents the total number of spatial features.

[0064] The spatial trajectory analysis algorithm consists of two core components: density clustering and trajectory segmentation. In cross-regional suspect tracking, the algorithm can identify the movement patterns of suspects across multiple cities and mark potential areas or hiding places.

[0065] Step 1.4: For the behavior dimension features, use the semantic tagging algorithm to add behavior tags to the trajectory points, focus on identifying suspicious behavior patterns, and form a behavior feature vector , the behavior feature vector is calculated as follows:

[0066]

[0067] in is the behavior feature extraction function, represents the original trajectory data, For the behavioral characteristics, Indicates the total number of behavioral characteristics.

[0068] The semantic tagging algorithm uses a multi-level behavior classification framework and combines contextual information to perform fine-grained annotation of behaviors. In the surveillance system, the algorithm can distinguish between normal activities of suspects (such as shopping and dining) and suspicious behaviors (such as repeated probing and following others), thereby improving the accuracy of behavior analysis.

[0069] Step 2: construct a temporal dependency graph of trajectory elements based on the feature set. The temporal dependency graph of trajectory elements is represented as a directed acyclic graph to capture the temporal relationship between trajectory elements.

[0070] This step constructs a temporal dependency graph of trajectory elements based on the extracted feature set to capture the temporal relationship between elements:

[0071] Step 2.1: Set the feature vectors extracted in step 1 Combined into a trajectory feature set ,in, represents a set of trajectory features, , , Respectively represent the 1st, 2nd, Trajectory elements, Represents the number of trajectory elements, and the trajectory elements are represented as time feature vectors, spatial feature vectors, or behavioral feature vectors.

[0072] The combination of trajectory elements adopts the feature fusion method, and the attention mechanism is used to give dynamic weights to features of different dimensions. When tracking suspects in serial cases, this method can automatically adjust the attention paid to time, space or behavior features according to the characteristics of the case. For example, for methods with strong temporal regularity, the weight of time dimension features will be increased.

[0073] Step 2.2, based on trajectory feature set , construct a directed acyclic graph ,in is the trajectory element set, is the temporal dependency edge set between elements, is the edge weight matrix; the construction of time-dependent edges adopts the time-dependent correlation algorithm to analyze the time sequence relationship and transition probability between elements and generate edge sets and the weight matrix ;

[0074] The time series correlation algorithm includes two key modules: conditional probability estimation and time interval analysis. In behavior prediction, the algorithm can learn the suspect's behavior sequence pattern, such as a series of pre-actions such as scouting, observation, and probing before committing a burglary.

[0075] Step 2.3: Apply the graph optimization algorithm to the constructed timing dependency graph for sparse processing, remove the edges with low correlation, and generate the optimized timing dependency graph. The sparseness process is based on the following formula:

[0076]

[0077] in Indicates that from the element To feature The directed edge of represents the set of temporal dependency edges between optimized elements, is the temporal dependency edge set between elements, represents the edge weight, is a threshold parameter used to control the sparsity of the graph.

[0078] The graph optimization algorithm uses an adaptive threshold strategy to dynamically adjust the sparsification degree according to the overall connection density of the graph. In large-scale suspect network analysis, the algorithm can retain key associations while removing noise connections and optimizing computing resource utilization.

[0079] Step 2.4, using the incremental graph update algorithm, when there are new trajectory elements When adding, only the temporal dependency relationship between the new element and the existing element is calculated, and the updated temporal dependency graph is generated by locally adjusting the updated graph structure. The computational complexity of the incremental update algorithm is , not the entire image is reconstructed ,in Indicates the size of the track feature collection.

[0080] The incremental graph update algorithm uses partition indexing technology to partition the nodes in the graph according to spatiotemporal features, and only calculates the relationship between the new node and the nodes in the relevant partition. In real-time monitoring scenarios, when new activity data of the suspect is received, the algorithm can quickly update the trajectory archive, allowing the monitoring system to reflect the latest behavior pattern changes within seconds.

[0081] Step 3: Using the constructed trajectory element time-series dependency graph, a trajectory dynamic system model based on Neural ODEs is constructed, expressed as a continuous-time dynamic system, to achieve trajectory state evolution modeling; the Neural ODE trajectory dynamic system model is used to achieve trajectory state inference at any time point and supports the recovery of continuous trajectories from finite observation points;

[0082] This step establishes a trajectory dynamic system model based on the Neural Ordinary Differential Equation to achieve trajectory state expression and inference in the continuous time domain:

[0083] Step 3.1, use the neural network to parameterize the derivative function of the ordinary differential equation and model the evolution of the trajectory state as a continuous-time dynamic system, expressed as:

[0084]

[0085] in for Dimension time The trajectory state vector, The parameters are The neural network, represents the rate of change of the trajectory state vector over time, is the dimension of the state vector.

[0086] Neural Networks The residual network architecture is adopted, which includes multiple layers of fully connected layers and activation functions. The network can capture complex nonlinear dynamic characteristics in suspect trajectory prediction, such as abnormal behavior patterns such as sudden turns, acceleration or deceleration, which has important application value in predicting the possible escape route of fugitives.

[0087] Step 3.2: Build and train the neural ODE model, use the features extracted in steps 1 and 2 and the time-dependent information as training data, and optimize the neural network parameters through back propagation. , so that the model can accurately express the evolution law of trajectory state. The training objective function is:

[0088]

[0089] in, represents the loss function; Represents the time predicted by the model Status; Indicates the actual observed time Status; represents the square of the Euclidean norm; Indicates the number of training samples; Representation parameters The regularization term of ; Represents the regularization coefficient, which is used to control the regularization strength.

[0090] The training of the neural ODE model uses an adaptive learning rate strategy and batch normalization technology to improve training stability. When tracking suspects in complex urban environments, the model can learn the constraints of the urban road network and the individual mobility preferences of suspects, so as to accurately predict their possible activity trajectories even in the blind spots of the camera.

[0091] Step 3.3, construct a hidden state inference algorithm, solve the differential equation through numerical integration, and realize the trajectory state inference at any time point. For a given initial state , inferred time Status The calculation is:

[0092]

[0093] in Initial time The trajectory state vector, To infer time The trajectory state vector, The parameters are The neural network, Indicates from time arrive 's points.

[0094] This integral is calculated using a numerical solver (such as the Runge-Kutta method with adaptive step size) to achieve high-precision state inference.

[0095] The hidden state inference algorithm uses an adaptive step-size control mechanism to dynamically adjust the integration step size according to the state change rate. In the case of a large amount of missing data in the suspect's trajectory, the algorithm can infer the complete trajectory based on known observation points, fill in the monitoring blind spot data, and support full-time behavior analysis.

[0096] Step 3.4: Construct a context-conditioned dynamics modulation system based on the attention mechanism, according to the environmental context information Dynamically adjust the neural ODE model to adapt to the trajectory evolution law under different environments. The formula for conditional dynamics modulation is:

[0097]

[0098] in, Represents the state vector About time The derivative of The parameters are Neural network; Indicates time The state vector of represents the time variable; Represents the contextual condition vector, which contains environmental characteristics such as weather, time period, area type, etc.

[0099] The context-conditioned dynamics modulation system processes context information from different sources through a multi-head attention mechanism. In the analysis of the impact of weather changes on suspect behavior, the system can capture changes in activity patterns in rainy and snowy weather and improve the accuracy of trajectory prediction in extreme environments.

[0100] Step 4: construct a continuous-discrete hybrid archive structure, save discrete observation points and continuous dynamics models at the same time, apply adaptive compression storage strategy, and build a multi-level index system;

[0101] This step builds a hybrid archive structure that contains both discrete observational data and a continuous kinetic model:

[0102] Step 4.1: Combine the trajectory element set extracted in step 1, the timing dependency graph constructed in step 2, and the dynamic system model obtained in step 3 into a unified archive structure. ,in is the trajectory element set, is the timing dependency graph, is a parameterized neural ODE model.

[0103] The hybrid archive structure adopts a hierarchical storage architecture to organize different types of data under a unified index system. In cross-departmental collaborative investigations, this structure can support various professionals to analyze the trajectory data of the same suspect from different perspectives, such as behavioral analysts focusing on behavioral patterns, spatial analysts focusing on geographical distribution, and time series analysts focusing on time patterns.

[0104] Step 4.2: Apply the adaptive compression storage algorithm to analyze the feature set based on information entropy and criticality. Compress, focus on retaining key nodes and behavioral features, and generate a compressed feature set The compression strategy is based on:

[0105]

[0106] in, Represents a compressed feature set; Represents the original feature set; Represents the first elements; Representation elements The information measurement function is used to evaluate the importance of factors; Indicates a dynamic threshold, which is adaptively adjusted by the system based on storage capacity and information preservation requirements.

[0107] The adaptive compression storage algorithm uses an information entropy weighted method to evaluate the importance of elements. In long-term monitoring data management, the algorithm can identify and retain complete information on key turning points (such as first entry into a sensitive area, abnormal stay behavior), while efficiently compressing routine activities to achieve long-term storage of massive data.

[0108] Step 4.3, build a multi-level index system, including structural index based on time-series dependency graph and semantic index based on feature vector, to support multi-granularity query requirements. Index structure includes: Time-series index: based on time-series dependency graph Build, support time series pattern retrieval - spatial index: based on spatial feature vector Build, support spatial region retrieval - behavior index: based on behavior feature vector Build,support specific behavior pattern retrieval - comprehensive index: integrate the above indexes, and support joint retrieval of complex conditions.

[0109] The multi-level index system uses a hierarchical inverted index structure to support multi-dimensional queries from macro to micro. In complex case investigations, investigators can quickly locate matching suspect trajectory fragments based on compound conditions such as being in a specific area, within a specific time period, and having brief contact with multiple people, greatly improving the efficiency of clue screening. The index system especially optimizes the performance of cross-dimensional queries, such as the spatiotemporal combination query of appearing at location A first and then appearing at location B within 24 hours.

[0110] Step 4.4, construct the trajectory archive query interface and restoration algorithm to support the restoration of complete trajectory information from the stored compressed elements and dynamic system models. The query operation quickly locates relevant elements based on the index system, supplements the state information of missing time points through the neural ODE model, and generates a continuous and complete trajectory expression.

[0111] The trajectory archive query interface adopts a hierarchical progressive query strategy, first performing efficient coarse-grained screening, and then fine-tuning the candidate results. In emergency response, the interface can identify suspicious persons who meet specific behavior patterns from millions of trajectory data within milliseconds, and generate complete activity trajectory reconstruction to provide support for rapid decision-making. The restoration algorithm supports multiple output formats, including visual trajectories, time series event lists, and statistical analysis reports to meet the needs of different usage scenarios.

[0112] In one embodiment of the present invention, an application example of the aforementioned trajectory archive construction method is provided;

[0113] The present invention focuses on the analysis of suspect trajectories in a certain area; suspect trajectory files need to support a variety of investigation scenarios, including: quickly locating the suspect's historical activity area, predicting the suspect's possible appearance location, identifying the suspect's abnormal activity pattern, and associating the common activity trajectories of multiple suspects;

[0114] Before applying the method of the present invention, the system collected the original trajectory data of the five target suspects within 30 days, including the appearance records captured by the camera, the electronic fence triggering records, and the mobile phone signal tower positioning records, etc. As shown in Table 1, the basic situation of the original trajectory data is shown;

[0115] Table 1: Statistics of original trajectory data

[0116]

[0117] The system cleans and aligns the raw data, removes obviously incorrect positioning points and records with abnormal timestamps, and finally obtains a trajectory dataset that can be used for subsequent analysis;

[0118] For the trajectory data of suspect S001, the system uses a multi-dimensional feature decomposition algorithm to process it and extract the time dimension, space dimension and behavior dimension features respectively. The feature extraction results are shown in Table 2:

[0119] Table 2: Multi-dimensional feature extraction results for suspect S001

[0120]

[0121] The system generates a hierarchical feature representation by fusing the extracted three-dimensional features. In particular, for suspect S001's activities in area A, the system identified an obvious spatiotemporal behavior pattern: every Tuesday and Friday between 22:00 and 24:00, the stay time is 40-50 minutes, and after leaving area A, there is an 80% probability of going to area D. This pattern is highly correlated with drug trafficking activities in area A.

[0122] Based on the extracted multi-dimensional features, the system constructed trajectory element temporal dependency graphs for the five suspects. Taking suspect S001 as an example, the system identified 26 key trajectory element nodes and constructed the temporal dependency relationships between them. As shown in Table 3, the temporal transition probabilities of some key element nodes are shown:

[0123] Table 3: Temporal transition probability matrix of some trajectory element nodes of suspect S001

[0124]

[0125] The initial timing dependency graph was sparsified using a graph optimization algorithm. The system removed edges with a transfer probability lower than a threshold value θ=0.1, and obtained an optimized timing dependency graph. The number of edges was reduced from 172 to 68, a decrease of 60.5%, while retaining the key transfer patterns.

[0126] In terms of applying the incremental graph update algorithm, when a new element v027 of suspect S001 in area A appears in the new observation data, the system only calculates the relationship between this element and the related element nodes of area A and area D, without rebuilding the entire graph structure. The update calculation time is reduced from the original 1,250 milliseconds to 125 milliseconds, and the performance is improved by about 90%.

[0127] The system uses a three-layer residual network structure parameterization Function, the input dimension is the state vector , including information such as position, speed, acceleration, and behavior pattern probability. 80% of the observation data within 30 days is used for training, and 10% of the data is used for validation and testing respectively. As shown in Table 4, the key parameters of model training are shown:

[0128] Table 4: Neural ODE model training parameters

[0129]

[0130] After training, the system predicts the key test segments in the suspect S001 trajectory. Specifically, given the initial state at time t=0 , predict the suspect's trajectory status in the next 8 hours , t∈[0,8].

[0131] In terms of prediction accuracy, as shown in Table 5, the performance of the neural ordinary differential equation model of the present invention and the traditional model on 50 test trajectory segments are compared:

[0132] Table 5: Comparison of prediction accuracy of different models (mean error ± standard deviation)

[0133]

[0134] The results show that the neural ordinary differential equation model of the present invention is superior to traditional methods in three aspects: position prediction, time prediction and behavior prediction. Especially in position prediction, it is improved by 57.6% compared with LSTM, and in time prediction, it is improved by 51.9%.

[0135] Based on the above multi-dimensional feature decomposition and dynamic system modeling results, the system builds trajectory files for the five target suspects. Taking suspect S001 as an example, its trajectory file contains a 4-level index structure, as shown in Table 6:

[0136] Table 6: Suspect S001 trajectory file index structure

[0137]

[0138] In terms of applying the continuous-discrete hybrid archive structure, the system stores discrete observation data points (a total of 15,382) and continuous trajectory function parameters generated by the Neural Ordinary Differential Equation. This hybrid structure reduces the original data storage requirements from 98.7MB to 12.3MB, reducing the storage space by 87.5%.

[0139] In terms of trajectory information retrieval, the following key query tests were conducted:

[0140] 1. Spatiotemporal joint retrieval: Query all activities of suspect S001 in area A from 2023-05-01 to 2023-05-07. The system returns 7 records and the retrieval time is 18 milliseconds.

[0141] 2. Behavior pattern retrieval: Query all instances of suspect S001 moving from area A to area D at night. The system returns 12 records and the retrieval time is 32 milliseconds.

[0142] 3. Time series path retrieval: query all trajectory segments that satisfy the path pattern v001→v008→v017. The system returns 3 records and the retrieval time is 45 milliseconds.

[0143] 4. Joint search of multiple suspects: Query the time period when S001 and S003 appeared together in area A on 2023-05-15. The system returns 1 record and the search time is 27 milliseconds.

[0144] Compared with the traditional relational database retrieval method, the multi-level index structure of the present invention has obvious advantages in complex queries, as shown in Table 7:

[0145] Table 7: Performance comparison of different retrieval methods on complex queries

[0146]

[0147] Through the above application examples, the trajectory file construction method of the present invention has demonstrated certain technical effects in actual suspect trajectory analysis:

[0148] Improved prediction accuracy: By using the Neural Ordinary Differential Equation model to build a dynamic system, the position prediction error was reduced to 78.6 meters, an improvement of 57.6% compared to the traditional LSTM method; the time prediction error was reduced to 12.3 minutes, an improvement of 51.9%; and the behavior prediction accuracy was increased to 92.1%, an improvement of 20.7%.

[0149] Improved computing efficiency:

[0150] In the construction of the time-dependent graph, the incremental update algorithm reduces the graph update calculation time by 90%;

[0151] The graph optimization algorithm reduces the number of edges by 60.5% while preserving the key transition patterns;

[0152] The continuous-discrete hybrid archive structure reduces storage space requirements by 87.5%;

[0153] Improved retrieval efficiency: In complex query scenarios, the multi-level index structure improves retrieval speed by an average of 58.1% compared to the Neo4j graph database and 79.6% compared to the MySQL relational database.

[0154] Overall effect of the system: In actual criminal investigation work, this system was used to analyze the trajectories of five target suspects, successfully identifying 12 important behavior patterns and seven key activity areas, providing valuable clues for the case-handling units and assisting in solving two related cases.

[0155] The above application examples and technical effect verifications show that the trajectory archive construction method of the present invention has certain advantages in prediction accuracy, calculation efficiency and retrieval efficiency, and provides an efficient and reliable technical means for suspect trajectory analysis.

[0156] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. A method for constructing a trajectory archive, characterized in that: The following steps are involved: Obtain the original trajectory data, perform multi-dimensional feature decomposition on the original trajectory data, and generate a time dimension feature set, a space dimension feature set, and a behavior dimension feature set; Based on the feature set, a temporal dependency graph of trajectory elements is constructed. The temporal dependency graph of trajectory elements is represented as a directed acyclic graph to capture the temporal relationship between trajectory elements. Using the constructed trajectory element time-series dependency graph, a trajectory dynamic system model based on Neural ODEs is constructed, expressed as a continuous-time dynamic system, to achieve trajectory state evolution modeling; the Neural ODE trajectory dynamic system model is used to achieve trajectory state inference at any time point, and supports the recovery of continuous trajectories from finite observation points; Construct a continuous-discrete hybrid archive structure to simultaneously save discrete observation points and continuous dynamic models, apply an adaptive compression storage strategy, and build a multi-level indexing system.

2. A method for constructing a trajectory file according to claim 1, characterized in that: In the multi-dimensional feature decomposition step, time dimension feature extraction includes: Apply the time segmentation algorithm to extract time series features, including frequency features, duration features, and periodic features, to form a time feature vector , which is calculated as: ； in is the time feature extraction function, represents the original trajectory data, For the time characteristics, Represents the total number of temporal features.

3. A method for constructing a trajectory file according to claim 1, characterized in that: In the multi-dimensional feature decomposition step, the spatial dimension feature extraction includes: using a spatial trajectory analysis algorithm to extract spatial pattern features from the trajectory point sequence, including the stop point, the moving path and the activity area, to form a spatial feature vector , which is calculated as: ； in is the spatial feature extraction function, represents the original trajectory data, For the spatial features, Represents the total number of spatial features.

4. A method for constructing a trajectory archive according to claim 1, characterized in that: In the multi-dimensional feature decomposition step, the behavior dimension feature extraction includes: using a semantic tagging algorithm to add behavior tags to the trajectory points, identifying suspicious behavior patterns, and forming a behavior feature vector. , which is calculated as: ； in is the behavior feature extraction function, represents the original trajectory data, For the behavioral characteristics, Indicates the total number of behavioral characteristics.

5. A method for constructing a trajectory archive according to claim 1, characterized in that: The step of constructing the trajectory element temporal dependency graph includes: combining the feature vector set into a trajectory element set: ； in, represents a set of trajectory features, , , Respectively represent the 1st, 2nd, Trajectory elements, represents the number of trajectory elements, where the trajectory elements are represented as time feature vectors, space feature vectors or behavior feature vectors; Constructing a directed acyclic graph ;in is the trajectory element set, is the temporal dependency edge set between elements, is the edge weight matrix; apply the graph optimization algorithm to perform sparse processing, remove low-correlation edges, and generate an optimized timing dependency graph ;in is the trajectory element set, is the set of temporal dependency edges between optimized elements, is the optimized edge weight matrix; An incremental graph update algorithm is adopted. When new trajectory elements are added, the graph structure is updated through local adjustments.

6. A method for constructing a trajectory archive according to claim 5, characterized in that: The sparsification process of the graph optimization algorithm is based on the following formula: ； in Indicates that from the element To feature The directed edge of represents the set of temporal dependency edges between optimized elements, is the temporal dependency edge set between elements, represents the edge weight, is the threshold parameter.

7. A method for constructing a trajectory archive according to claim 1, characterized in that: The trajectory dynamic system model of the Neural Ordinary Differential Equation is expressed as: ； in for Dimension time The trajectory state vector, The parameters are The neural network, represents the rate of change of the trajectory state vector over time, is the dimension of the state vector.

8. A method for constructing a trajectory archive according to claim 1, characterized in that: The trajectory dynamic system model of the Neural Ordinary Differential Equation includes a hidden state inference algorithm, which realizes trajectory state inference at any time point by solving the differential equation through numerical integration: ； in Initial time The trajectory state vector, To infer time The trajectory state vector, The parameters are The neural network, Indicates from time arrive 's points.

9. A method for constructing a trajectory archive according to claim 1, characterized in that: The continuous-discrete hybrid archive structure construction step comprises: Combine the collection of trajectory elements, the timing dependency graph, and the dynamic system model into a unified archive structure: ； Apply adaptive compression storage algorithm to compress feature collection, focusing on retaining key nodes and behavioral features; Build a multi-level index system, including time index, space index, behavior pattern index and time-dependent path index; Build trajectory archive query interface and restoration algorithm to support restoration of complete trajectory information from stored compressed elements and dynamic system models; in, Represents the file structure, is the trajectory element set, is the optimized timing dependency graph. The parameters are Dynamic system model.

10. A trajectory file construction system, used to execute a trajectory file construction method according to any one of claims 1 to 9, characterized in that: include: A multi-dimensional feature decomposition module is used to perform multi-dimensional feature decomposition on the original trajectory data to generate a time dimension feature set, a space dimension feature set, and a behavior dimension feature set; The timing dependency graph construction module is used to construct the trajectory element timing dependency graph. The trajectory element timing dependency graph is represented as a directed acyclic graph to capture the timing relationship between trajectory elements. The Neural Ordinary Differential Equation Model module is used to construct the trajectory dynamic system model, expressed as a continuous-time dynamic system, and realize the evolution modeling of trajectory states; The hybrid archive structure module is used to construct a continuous-discrete hybrid archive structure, save discrete observation points and continuous dynamic models at the same time, apply adaptive compression storage strategies, and build a multi-level index system.

Images