An online trajectory stitching method and system for real-time data streams

Abstract

The application discloses an online trajectory connection method for real-time data flow, and comprises the following steps: first, a historical target management library is established; second, real-time input trajectory data flow is received, a dynamic growing trajectory is constructed and updated every time a new trajectory point is received; third, when the target identity of the dynamic growing trajectory is not confirmed, a new trajectory point is added and the trajectory is updated, and then a historical target meeting preset candidate conditions is screened to form a candidate target set; fourth, online association reasoning is performed on the dynamic growing trajectory and each candidate historical target in the candidate target set, and an online association result is output; fifth, a sequence of association results is obtained; sixth, identity judgment is performed according to continuous results in the sequence of association results; and finally, trajectory coding features, physical motion information and trajectory cache data are updated based on the complete trajectory point sequence of the dynamic growing trajectory.

Description

Technical Field

[0001] This invention relates to the field of maritime target monitoring and tracking technology, specifically an online trajectory continuation method and system for real-time data streams. Background Technology

[0002] With the continuous development of intelligent shipping, maritime traffic monitoring, and continuous target tracking technologies, various sensing devices such as Automatic Identification Systems (AIS) and radar are widely used in maritime target monitoring tasks. These devices can continuously acquire target position, time, and other motion information, and form target trajectories based on continuous observation data, providing a data foundation for ship monitoring, trajectory prediction, and target situation analysis. However, in practical applications, due to factors such as limited sensor coverage and complex and variable communication environments, continuous observation data for the same target is prone to temporal discontinuities or spatial gaps, causing the complete target trajectory to be divided into multiple independent trajectory segments. Therefore, it is necessary to determine whether different trajectory segments belong to the same real target and establish the continuity relationship between trajectory segments to reconstruct the continuous motion process of the target.

[0003] Currently, existing methods for trajectory continuation mainly include trajectory matching methods based on motion constraints, association methods based on probabilistic models, and trajectory association methods based on deep learning. Among these, methods based on motion constraints typically utilize information such as time intervals, spatial distances, speed differences, and heading changes between trajectories to determine the association relationship between them through preset rules or thresholds. Methods based on motion models and statistical inference calculate the trajectory association probability by establishing a target motion model. Deep learning-based methods use neural networks to extract the spatiotemporal features of trajectories and determine whether different trajectories belong to the same target based on the model output. These methods can achieve good association results when the trajectory length is sufficient and the motion information is complete.

[0004] However, the existing methods described above still have the following shortcomings: In practical maritime traffic surveillance and continuous target tracking missions, trajectory data typically arrives continuously in the form of a data stream. The target trajectory is in a dynamic growth process from the moment the first trajectory point is received, and its motion state gradually becomes apparent over time. Target identity usually needs to be confirmed as early as possible before trajectory reception is complete to meet the timeliness requirements of tasks such as real-time early warning and situation assessment. However, existing methods all require waiting for the complete trajectory segment to form before association judgments are made. For example, motion constraint-based methods need to calculate statistics such as speed and heading of the complete trajectory segment; probabilistic model-based methods require a complete trajectory sequence to estimate motion model parameters; and deep learning-based methods also use complete trajectory segments as input for association classification. This offline processing mode, which relies on complete trajectories, means that association results can only be output once after trajectory reception, and cannot be gradually provided as the trajectory continues to grow, severely limiting the response speed in online application scenarios.

[0005] Existing methods treat trajectory association as a one-time matching process, making an association decision based solely on the similarity or probability between a complete trajectory segment and a candidate historical target. This single-judgment model suffers from significant fluctuations in the reliability of single-association inference results when faced with insufficient initial trajectory motion information, high observation noise, or dense target distribution. It is prone to erroneous association confirmations due to random noise or insufficient information. Furthermore, existing methods lack a dynamic evaluation mechanism for the reliability of association results. They cannot verify the stability of association conclusions through continuous observation, nor can they gradually accumulate confidence and progressively determine target identity using multiple consecutive association outputs. Once a single judgment is incorrect, the erroneous association result will be fixed in subsequent processes, leading to irreversible misidentification.

[0006] In practical applications of some existing methods, once it is confirmed that the current trajectory belongs to the same identity as a historical target, the motion characteristics and cached data of that historical target are updated using the currently received trajectory points to better reflect the target's latest motion state. However, when the target's identity is confirmed early in the trajectory reception process, the current trajectory contains only a small number of trajectory points, its motion characteristics are still unstable, and the trajectory shape is not yet fully revealed. Updating historical target information using such an incomplete trajectory will cause a shift in the reference features in the historical target management database. This shift will directly affect the historical references used by other unconfirmed dynamically growing trajectories in subsequent candidate target screening and association reasoning, causing previously correct association judgments to become unstable due to changes in references, and even triggering a cascading effect of multi-target identity confusion. Existing offline association methods do not have this problem because they only make judgments after the trajectory is fully formed; however, for online scenarios that require early identity confirmation, how to achieve rapid identity confirmation while avoiding incomplete information contaminating historical references is a key problem that has not yet been effectively solved.

[0007] In summary, existing trajectory continuation methods mainly rely on offline, one-time association of complete trajectories, which is insufficient to meet the comprehensive requirements of real-time data stream scenarios for response timeliness, decision reliability, and process stability. There is an urgent need for an online trajectory continuation method that can progressively confirm the target identity during trajectory reception without interfering with subsequent association processes due to early confirmation. Summary of the Invention

[0008] To address the problems of existing trajectory continuation methods in the background art, which mainly rely on offline association of complete trajectories and are difficult to adapt to real-time data stream scenarios, this invention provides an online trajectory continuation method and system for real-time data streams. Based on the results of multiple consecutive online associations, the target identity is confirmed before the trajectory reception is completed. After the identity is confirmed, trajectory points are received until the complete trajectory is formed, and then the historical target management database is updated uniformly. This balances the real-time nature of identity confirmation, the stability of association results, and the consistency of historical references.

[0009] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows: An online trajectory continuation method for real-time data streams includes the following steps: Step 1: Establish a historical target management database, assign a unique target identification number to each historical target, and save the historical trajectory coding features, historical physical motion information, and trajectory cache data generated from the most recently received complete trajectory segment of each historical target. Step 2: Receive the real-time input trajectory data stream. For each new trajectory point received, construct and update the dynamically growing trajectory based on the real-time trajectory identifier carried by the trajectory point. Step 3: When the target identity of the dynamically growing trajectory has not yet been confirmed, whenever a new trajectory point is added and the trajectory update is completed, based on the updated current physical motion information and the historical physical motion information of each historical target in the historical target management database, historical targets that meet the preset candidate conditions are selected to form a candidate target set. Step 4: When the target identity of the dynamically growing trajectory has not yet been confirmed, whenever a new trajectory point is added and the candidate target set is updated, online association reasoning is performed on the dynamically growing trajectory and each candidate historical target in the candidate target set, and the online association results are output. Step 5: Add each online association result to the association result sequence of the corresponding dynamic growth trajectory in the order of generation; perform identity judgment based on the continuous results in the association result sequence; after determining the target identity based on the judgment result, lock the association information and continue to receive subsequent trajectory points until the trajectory reception is completed; Step 6: When the reception status of the dynamically growing trajectory is "reception completed" and the association status is "confirmed", generate trajectory coding features, physical motion information and trajectory cache data based on the complete trajectory point sequence of the dynamically growing trajectory. Update the corresponding historical target record or create a new historical target record according to the confirmed association information. Use the updated historical target record as a historical reference for subsequent real-time trajectory continuation, so that the trajectory segments formed by the same target can continue to be associated with the corresponding target identity, so as to maintain the continuity of target identity and trajectory information.

[0010] Preferably, in step 2, the method for constructing and updating the dynamic growth trajectory is as follows: find the corresponding dynamic growth trajectory; if there is no corresponding trajectory, create a new dynamic growth trajectory and use the trajectory point as the first trajectory point; if there is already a corresponding trajectory, add the trajectory point to the trajectory point sequence; after each addition of a new trajectory point, update the physical motion information of the current trajectory synchronously according to the updated trajectory point sequence; each dynamic growth trajectory carries at least a real-time trajectory identifier, a trajectory point sequence, current physical motion information, an association status indicating whether the target identity has been confirmed, association information, and a trajectory reception status indicating whether the trajectory reception is complete.

[0011] Preferably, in step 3, the preset candidate conditions include time interval constraints, spatial distance constraints, and motion consistency constraints; wherein: The time interval constraint is: the time interval between the start time of the dynamic growth trajectory and the end time of the most recent complete trajectory segment of the historical target does not exceed a preset time threshold. The spatial distance constraint is that the distance between the starting position of the dynamic growth trajectory and the ending position of the most recent complete trajectory segment of the historical target does not exceed a preset distance threshold. The motion consistency constraint is as follows: when the corresponding motion parameters are available, the difference between the velocity of the dynamic growth trajectory and the velocity of the historical target does not exceed a preset velocity difference threshold, and the difference between the heading angle of the dynamic growth trajectory and the heading angle of the historical target does not exceed a preset heading difference threshold; when the corresponding motion parameters are unavailable, the corresponding motion consistency constraint is not executed. Historical targets that simultaneously satisfy the time interval constraint, the spatial distance constraint, and the motion consistency constraint are added to the candidate target set.

[0012] Preferably, the method for online correlation reasoning in step 4 is as follows: The current trajectory encoding features are generated based on the updated trajectory point sequence. The historical trajectory encoding features of candidate historical targets are read from the historical target management library, and the physical motion difference features are calculated based on the current physical motion information and the historical physical motion information of candidate historical targets. The current trajectory encoding features, the historical trajectory encoding features, and the physical motion difference features are fused and input into the trajectory continuation neural network model to obtain the predicted probability that the current dynamically growing trajectory and the candidate historical target belong to the same target, and this predicted probability is used as the association confidence. The candidate historical target with the highest association confidence is selected as the best association target for this time, and the online association result for this time is output.

[0013] Preferably, in step 4, the trajectory continuation neural network model includes a trajectory feature extraction network, a physical motion feature extraction network, and a fusion classification network; The trajectory feature extraction network includes a self-supervised pre-trained trajectory encoder, which generates current trajectory encoding features based on the trajectory feature tensor composed of lateral position, longitudinal position, lateral position change, and longitudinal position change, effective point mask, and actual trajectory length. The current trajectory encoding features are then combined with the historical trajectory encoding features of candidate historical targets read from the historical target management database, as well as the absolute difference features and product features of the two to obtain trajectory matching features. The physical motion feature extraction network is used to extract physical motion features under a unified local relative coordinate system based on the trajectory cache data of candidate historical targets, historical physical motion information, and the position coordinate sequence and sampling time of the current dynamically growing trajectory, and to generate physical embedding features. The fusion classification network is used to concatenate the trajectory matching features with the physical embedding features to form fusion features. After passing through a binary classification output layer and a Softmax function, the class probabilities corresponding to the trajectory discontinuity category and the trajectory continuity category are obtained, and the probability corresponding to the trajectory continuity category is used as the association confidence.

[0014] Preferably, in step 5, the method for determining identity based on consecutive results in the association result sequence is as follows: when multiple consecutive association results meet the preset stable association conditions, the association information of the current dynamic growth trajectory is set as the target identity number of the corresponding historical target, and the association status is updated to confirmed; when the preset independent target condition is met, the association information of the current dynamic growth trajectory is marked as a new target, and the association status is updated to confirmed; when none of the above conditions are met, the association status is kept as unconfirmed, and steps 2 to 5 are continued after a new trajectory point is added; after the target identity is determined, the association information is locked, the candidate target screening and online association reasoning for identity selection are stopped for the dynamic growth trajectory, and subsequent trajectory points are received until the trajectory reception is completed.

[0015] Preferably, in step 5, the stable association condition includes a target number consistency condition and an association credibility stability condition; a judgment window length and an association credibility threshold are set. If the best associated target number remains consistent in the most recent consecutive association inference results of the judgment window length, and the association credibility is higher than the association credibility threshold each time within the window, then the stable association condition is determined to be satisfied.

[0016] Preferably, in step 5, the independent target condition is: set a judgment window length and an independent judgment threshold. If the association credibility of the most recent consecutive association reasoning results of the judgment window length is lower than the independent judgment threshold, then the current trajectory is determined not to belong to any existing historical target, thus satisfying the independent target condition. The zero association credibility result generated when the candidate target set is empty in step 4 participates in the continuous judgment in the same way.

[0017] Preferably, in step 5, when the dynamic growth trajectory is still in an unconfirmed state, and the number of trajectory points reaches the preset maximum inference points or the trajectory reception status changes to reception completed, a termination association judgment is performed; the termination association judgment is as follows: read the last online association result, and when the best candidate historical target is not empty and the corresponding association credibility is not lower than the preset termination association threshold, set the association information to the target identity number corresponding to the historical target; otherwise, mark the association information as a new target; and update the association status to confirmed.

[0018] Preferably, in step 5, when the trajectory meets the stable association condition, the time range of the dynamically growing trajectory is compared with the time range of other dynamically growing trajectories that have been associated with the same historical target. If the time ranges do not overlap, association with the same historical target is allowed. If the time ranges overlap, the association result of the previously confirmed trajectory is retained, and the historical target is excluded from the candidate set of the subsequently confirmed trajectory. Online association reasoning is then re-executed based on the currently received trajectory points. The result of the re-reasoning is added to the association result sequence as a valid association result, the original association result is discarded, and association judgment continues to be performed in the order of stable association condition, independent target condition, and termination of association judgment.

[0019] Preferably, in step 1, before the dynamic growth trajectory reception is completed, the historical trajectory encoding features, historical physical motion information and trajectory cache data of each historical target in the historical target management library are kept unchanged.

[0020] Preferably, in step 2, when a trajectory end marker is received or no new trajectory point is received after a preset time, the trajectory receiving status of the corresponding dynamically growing trajectory is updated to reception completed.

[0021] Preferably, in step 6, if the associated information is the target identity number of an existing historical target, the corresponding historical target record is updated using the generated information; if the associated information is a new target marker, a new target identity number is assigned to the dynamic growth trajectory, a new historical target record is established, and it is added to the historical target management database.

[0022] This invention also provides an online trajectory continuity system for real-time data streams, comprising: The historical target management module is used to establish a historical target management library, assign a unique target number to each historical target, and save the trajectory coding features, trajectory physical motion information, and trajectory cache data corresponding to the most recent complete continuous trajectory segment of each historical target; before the dynamic growth trajectory reception is completed, the trajectory coding features, trajectory physical motion information, and trajectory cache data of each historical target in the historical target management library remain unchanged; The dynamic growth trajectory management module is used to create or update the corresponding dynamic growth trajectory based on the real-time trajectory identifier carried by the trajectory point each time a new valid trajectory point is received, and to continue receiving subsequent trajectory points after identity verification until the trajectory reception is completed. The candidate target dynamic filtering module is used to re-filter candidate historical targets whenever the dynamic growth trajectory is updated due to the addition of new valid trajectory points, before the identity of the dynamic growth trajectory has been confirmed. The online trajectory association reasoning module is used to perform online association reasoning once for each candidate historical target and the dynamically growing trajectory whenever the candidate target set is updated, and output the best associated target and its association confidence based on the association confidence. The trajectory association status dynamic update module is used to add each online association result to the association result sequence and perform an association judgment once. Before the trajectory is received, the target identity is confirmed based on multiple consecutive online association results, and the association information is locked after the target identity is confirmed. The historical target update module is used to keep the characteristics of historical targets in the historical target management library unchanged before the dynamic growth trajectory is received, and to update the corresponding historical target record or create a new historical target record based on the complete trajectory after the dynamic growth trajectory is received. The results output module is used to output the target number, association status, and association confidence level corresponding to the dynamic growth trajectory.

[0023] Preferably, the online trajectory association reasoning module is equipped with a trajectory continuation neural network model, which includes a trajectory feature extraction network, a physical motion feature extraction network, and a fusion classification network. The trajectory feature extraction network includes a self-supervised pre-trained trajectory encoder, which generates current trajectory encoding features based on the trajectory feature tensor composed of lateral position, longitudinal position, lateral position change, and longitudinal position change, effective point mask, and actual trajectory length. The current trajectory encoding features are then combined with the historical trajectory encoding features of candidate historical targets read from the historical target management database, as well as the absolute difference features and product features of the two to obtain trajectory matching features. The physical motion feature extraction network is used to extract physical motion features under a unified local relative coordinate system based on the trajectory cache data of candidate historical targets, historical physical motion information, and the position coordinate sequence and sampling time of the current dynamically growing trajectory, and to generate physical embedding features. The fusion classification network is used to concatenate the trajectory matching features and the physical embedding features to form fusion features. After passing through a binary classification output layer and a Softmax function, the probability of the trajectory discontinuity category and the trajectory continuity category is obtained, and the probability of the trajectory continuity category is used as the association confidence.

[0024] This invention has the following characteristics and beneficial effects: This invention targets real-time data stream environments. By constructing dynamically growing trajectories and continuously triggering online association reasoning with the input of trajectory points, it achieves gradual confirmation and continuous output of trajectory association results. This avoids the reliance on complete trajectories for one-time association judgment in traditional offline methods, thus improving the continuity response speed in real-time data stream scenarios. Simultaneously, it introduces a dynamic update mechanism for trajectory association status based on multiple consecutive association reasoning results. Through dual judgment of target number consistency and association credibility stability within a sliding window, it reduces the risk of erroneous associations caused by insufficient initial trajectory information or single-inference deviations, improving the stability of online trajectory continuity. Furthermore, after target identity confirmation, it continues to receive complete trajectory segments and updates historical target information accordingly with a delay. This ensures the integrity and consistency of trajectory encoding features, physical motion information, and cached data in the historical target management database, providing a more reliable historical reference for subsequent trajectory continuity and forming a virtuous cycle between historical target management and online continuity processes. Attached Figure Description

[0025] Figure 1 This is a flowchart of the online trajectory continuation method according to an embodiment of the present invention; Figure 2 This is a block diagram of the online trajectory continuity system according to an embodiment of the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0027] This embodiment provides an online trajectory continuation method for real-time data streams, such as... Figure 1 As shown, in order to simulate the continuous arrival of real-time trajectory data using offline scene trajectory data, this embodiment first preprocesses the offline scene trajectory data and constructs a global event sequence according to the sampling time of the trajectory points to form a repeatable point-by-point input process.

[0028] In this embodiment, trajectory data is organized by scene, with each scene containing several continuous trajectory segments. Each trajectory segment is stored in CSV file format, with key fields including time (absolute seconds within a day), latitude and longitude coordinates, and longitude and latitude coordinates. , The data includes the data batch number and the data source number, which together constitute the unique trajectory identifier for the continuous trajectory. All trajectory points within the same continuous trajectory share the same... Different continuous trajectories pass through different Distinguishing between them. The aforementioned This is only used to identify the continuous trajectory segment to which the trajectory point belongs and does not represent the true identity of the target. To ensure that the experimental data can stably simulate real-time data streams, a basic integrity check is performed on each trajectory segment to remove records with missing or abnormal time fields, out-of-bounds latitude and longitude coordinates, or speed mutations exceeding preset thresholds. After removing abnormal records, the trajectory points within each trajectory segment are arranged in ascending order of time to form a standardized trajectory sequence. It should also be noted that the trajectory points within each trajectory segment are spaced at the same interval, which is 5 seconds in this embodiment. Subsequently, the standardized trajectory point sequences within the same scene are merged and globally sorted according to the sampling time. Each event Corresponding to an independent trajectory point and its The order of events in the global event sequence indicates the sequence in which trajectory points are input into the online trajectory continuation system. This sorting is used to convert offline scene data into a repeatable point-by-point input process; out-of-order, duplicate, or late trajectory points that occur during actual online operation are handled by subsequent dynamically growing trajectory management steps according to preset rules.

[0029] After completing the offline scene data preprocessing and global event sequence construction, the trajectory points are input point by point into the online trajectory continuation system according to the global event sequence to simulate the continuous arrival of actual real-time data streams. The online trajectory continuation method specifically includes the following steps: Step 1: Establish a historical target management library, assign a unique target identification number to each historical target, and save the historical trajectory encoding features, historical physical motion information, and trajectory cache data generated from the most recently received complete trajectory segment of each historical target.

[0030] Furthermore, this embodiment first establishes a historical target management database to uniformly manage historical target information for which trajectory reception has been completed. During initial system runtime, the historical target management database is initialized as an empty set; as the online trajectory continuation process continues, subsequent step 6 gradually completes the establishment and updating of historical target records after the dynamic growth trajectory reception is finished.

[0031] Specifically, the historical target management database does not store the entire historical trajectory of a target, but only the key information of the target's most recent complete and continuous trajectory segment. For each historical target to be stored in the management database, a unique identification number is first assigned to it. The target remains unchanged after creation. Subsequently, the trajectory encoder in the trajectory feature extraction network is used to extract temporal features from the trajectory segment, resulting in fixed-dimensional trajectory encoded features. This is used to characterize the overall motion pattern and temporal variation pattern of the target. Simultaneously, it extracts the physical motion information corresponding to the end time of the trajectory segment. Specifically, this includes the target position coordinates, velocity, heading angle, and end time at the end of the trajectory segment. Additionally, a preset number of trajectory points at the end of the trajectory segment are saved as trajectory cache data. In this embodiment, a preset number of 200 trajectory points is used to retain the trajectory details of the historical target in the most recent period. Before the current dynamically growing trajectory is received, even if the target identity has been determined in advance, the historical trajectory encoding characteristics, historical physical motion information, and trajectory cache data of the corresponding historical target remain unchanged, avoiding the use of incomplete trajectories to change the historical information on which subsequent online association is based.

[0032] Specifically, in the historical goals The calculation method is as follows: convert the latitude and longitude coordinates of the trajectory point into local planar coordinates. Let the trajectory point be... ,in, Sampling time, and These are latitude and longitude, respectively. Selecting the endpoint... The trajectory point with the shortest time among all trajectory points is used as the local reference point. After converting latitude and longitude from degrees to radians, calculate the local planar coordinates of the trajectory points using the following formula: in, The average radius of the Earth Indicates displacement in the east-west direction. This represents displacement in the north-south direction. Let the trajectory point be represented by the coordinate transformation as... For adjacent valid trajectory points, calculate the velocity components. Adjacent point pairs with a time difference not greater than zero or with abnormal displacement are excluded from the calculation. For historical targets, a threshold is set for calculating historical motion characteristics. When the number of valid points in the complete trajectory segment is less than When the speed and heading angle are set to unavailable; when the number of valid points is not less than When selecting the nearest endpoint of the trajectory Calculate the effective trajectory points and the average value of the velocity components for each adjacent point. In this embodiment =32. Calculate the velocity and heading angle based on the average velocity components. Among them, the heading angle is due north. Increases in a clockwise direction; when At that time, the target has no clear direction of movement, and the heading angle is set to unavailable.

[0033] Step 2: Construct a dynamic growth trajectory. Receive real-time input trajectory data streams. For each new valid trajectory point received, search for the corresponding dynamic growth trajectory based on the real-time trajectory identifier carried by that trajectory point. If no corresponding trajectory exists, create a new dynamic growth trajectory and use that trajectory point as the first trajectory point. If a corresponding trajectory already exists, add that trajectory point to the trajectory point sequence. After each new trajectory point is added, synchronously update the physical motion information of the current trajectory based on the updated trajectory point sequence. Each dynamic growth trajectory carries at least a real-time trajectory identifier, a trajectory point sequence, current physical motion information, an association status indicating whether the target identity has been confirmed, association information, and a trajectory reception status indicating whether the trajectory reception is complete.

[0034] Specifically, in this embodiment, trajectory data is received point by point according to the global event sequence. Each time a new trajectory point is received, its trajectory identifier is used... Search for the corresponding dynamic growth trajectory in the current set of dynamic growth trajectories.

[0035] When there is no identical When creating a dynamic growth trajectory, a new dynamic growth trajectory is created using the current trajectory point. The trajectory point is added to the trajectory point sequence, the target identity information is initialized to empty, the associated status is initialized to "identity verification in progress," and the receiving status is initialized to "receiving in progress." The created dynamic growth trajectory is then added to the dynamic growth trajectory set. When a corresponding dynamic growth trajectory exists, its receiving status is checked first. If the dynamic growth trajectory is already in the "receiving completed" state, the current trajectory point is considered an abnormally late point and is not added to the dynamic growth trajectory. If the dynamic growth trajectory is still in the "receiving in progress" state, it continues to determine whether the current trajectory point is a duplicate of an existing trajectory point in the trajectory point sequence.

[0036] Furthermore, based on the time and position coordinates of the trajectory point, it is determined whether the trajectory point is repeated in the sequence of trajectory points in the dynamically growing trajectory. Duplicate trajectory points are discarded directly without triggering subsequent candidate set filtering. For non-duplicate trajectory points, their time and physical motion information are compared. The end time is determined by the time of the current trajectory point. If the time of the current trajectory point is later than the end time, it is added to the end of the trajectory point sequence; if the timestamp of the current trajectory point is earlier than the timestamp of the nearest trajectory point, the trajectory point is determined to be an out-of-order arrival point and is inserted into the trajectory point sequence in ascending order of time.

[0037] Furthermore, the heading angle, velocity, end time, and starting point coordinates in the current physical motion information are recalculated based on the updated ordered trajectory point sequence. Simultaneously, the system records the arrival time of the last valid trajectory point for each dynamically growing trajectory and continuously monitors its reception status. The system will terminate the process if it receives a trajectory end marker from the data source, or if a preset timeout period has elapsed since the arrival of the last valid trajectory point without receiving a marker with the same information. When a non-repeating valid trajectory point is found, the reception status of the dynamically growing trajectory is updated to reception complete. In this embodiment, the preset timeout is set to 20 minutes.

[0038] Furthermore, in the dynamic growth trajectory The calculation method is as follows: using the initial point in the trajectory point sequence as a local reference point, the planar coordinates of each trajectory point are obtained according to the aforementioned local planar coordinate transformation method. A threshold for calculating dynamic motion features is set. When the number of valid points is less than When the speed and heading angle are set to unavailable; when the number of valid points is not less than At this time, all currently received valid trajectory points are used, and the speed and heading angle are calculated according to the aforementioned method. In this embodiment, =8.

[0039] Step 3: Dynamic screening of candidate targets. When the dynamic growth trajectory is in the identity verification stage, after each new valid trajectory point is added and the physical motion information is updated, based on the temporal continuity, spatial proximity, and motion consistency between the current dynamic growth trajectory and historical targets, historical targets that may form a successor relationship with it are screened from the historical target management library to form a set of candidate targets, so as to narrow the search scope of subsequent online association reasoning.

[0040] Specifically, the time interval constraint is executed first. Let the starting time of the current dynamic growth trajectory be... Historical goals The latest trajectory segment ended at time At this moment, the physical motion information of the trajectory can be obtained. The time interval between the two can be calculated directly from the data. .like If the historical target meets the time continuity requirement, it is considered to satisfy the requirement; otherwise, it is directly eliminated. In this embodiment, the time threshold... Take 1200 seconds.

[0041] Furthermore, for historical targets filtered by time constraints, their cached data is utilized. Spatial distance constraints are applied to the tail trajectory points stored in the database. Let the coordinates of the starting point of the dynamically growing trajectory be... The position of the last trajectory point in the historical target cache data is Calculate the Haversine spherical distance between the two: in, , , The average radius of the Earth is taken as 6,371,000 meters. If... If the distance is within a certain range, the spatial proximity condition is met; otherwise, it is discarded. In this embodiment, the spatial distance threshold is... Take 6000 meters.

[0042] For the remaining historical targets that have passed the above two constraints, further apply motion consistency constraints. This is done by analyzing the physical motion information of the current dynamically growing trajectory. and historical target physical motion information Obtain speed and heading angle, and calculate speed difference and heading difference. In this embodiment, the speed difference threshold Take 5 m / s as the heading difference threshold. Take 150°. When all velocity parameters are available, the following conditions must be met. When all heading angle parameters are available, the following conditions must be met: If all the above executable motion consistency constraints are satisfied, the historical target is retained; otherwise, it is discarded. When the velocity of either the historical target or the dynamic growth trajectory is empty, the velocity difference constraint is not executed; when the heading angle of either side is empty, the heading difference constraint is not executed; when both velocity and heading angle are empty, candidate targets are determined only based on time continuity constraints and spatial distance constraints. The corresponding motion consistency constraints are executed only after the number of points on the dynamic growth trajectory increases and an effective velocity or heading angle is formed. In this embodiment, the velocity difference threshold and heading difference threshold are set relatively wide to reduce trajectory noise and erroneous rejection caused by insufficient initial motion information, and the fine distinction between candidate targets is handled by the association model in step 4.

[0043] After the above three-level progressive screening, the dynamic growth trajectory is obtained. Corresponding candidate target set in This represents the number of candidate targets. Through this filtering mechanism, most targets in the historical target management database that do not meet the requirements of spatiotemporal continuity and motion consistency are eliminated in advance. The number of candidate targets is typically reduced from tens to hundreds to just a few, significantly reducing the computational scale of online trajectory association inference in step 4, while also caching data. The use of this avoids reading the complete historical trajectory every time filtering, effectively improving online processing efficiency.

[0044] Step 4: Online Trajectory Association Reasoning. In this embodiment, when the target identity of the dynamically growing trajectory has not yet been confirmed, whenever a new valid trajectory point is added and the candidate target set is updated, online association reasoning is performed once for each candidate historical target in the candidate target set and the dynamically growing trajectory. The current trajectory encoding features are generated based on the updated trajectory point sequence. The historical trajectory encoding features of the candidate historical targets are read from the historical target management library, and the physical motion difference features are calculated based on the current physical motion information and the historical physical motion information of the candidate historical targets. The current trajectory encoding features, the historical trajectory encoding features, and the physical motion difference features are fused and input into the trajectory continuation neural network model to obtain the predicted probability that the current dynamically growing trajectory and the candidate historical targets belong to the same target, and this predicted probability is used as the association confidence level. The candidate historical target with the highest association confidence level is selected as the best association target for this time. Output the results of this online association. : in, Indicates the current dynamic growth trajectory and the first The association credibility between candidate historical targets is determined. When the candidate target set is empty, model inference is not performed, and an empty target and zero association credibility are directly output. The best associated target and its association credibility constitute an online association result and are used for continuous association judgment in step 5.

[0045] Specifically, a trajectory continuation neural network model is constructed, which includes a trajectory feature extraction network, a physical motion feature extraction network, and a fusion classification network.

[0046] Furthermore, the trajectory feature extraction network is used to generate trajectory encoding features of historical targets and real-time encoding features of dynamically growing trajectories, and obtains trajectory matching features based on the two.

[0047] Specifically, based on the trajectory's position coordinates, the lateral and longitudinal position changes are calculated, constructing a four-channel trajectory feature tensor composed of lateral position, longitudinal position, lateral position change, and longitudinal position change. This trajectory feature tensor, the effective point mask, and the actual trajectory length are input into a self-supervised pre-trained trajectory encoder. After multi-channel temporal feature extraction, channel fusion, and pooling, fixed-dimensional trajectory encoding features are obtained. Specifically, when historical targets are established or updated, the complete trajectory segment corresponding to the historical target is input into the trajectory encoder to obtain historical trajectory encoding features, which are stored in a historical target management library. During online association reasoning, the historical trajectory encoding features of each candidate historical target are read from the historical target management library, and the effective trajectory points received by the current dynamically growing trajectory are input into the trajectory encoder to obtain the current trajectory encoding features. The historical trajectory encoding features, the current trajectory encoding features, and their absolute difference and product features are combined, and after feature mapping, a 64-dimensional trajectory matching feature is obtained.

[0048] Furthermore, the physical motion feature extraction network takes the trajectory cache data of candidate historical targets, physical motion information, and the position coordinate sequence and sampling time of the current dynamically growing trajectory as input. It converts the end trajectory points of the candidate historical targets and the current dynamically growing trajectory into unified local relative coordinates, using the position of the end of the previous trajectory as the common relative coordinate origin. Then, it extracts physical motion features such as differences in motion trends to form an original physical feature vector. After normalization and mapping by a fully connected network, the original physical feature vector outputs a 64-dimensional physical embedding feature.

[0049] For any preset trajectory length scale, the least squares method is used to linearly fit the lateral and longitudinal relative coordinates of the preceding and subsequent trajectories to the sampling time, and the difference in motion trend is calculated based on the difference in the corresponding fitting slopes. In the formula, This indicates the preset trajectory length scale. and These represent the lateral and longitudinal fitting slopes of the preceding trajectory at this scale, respectively. and These represent the lateral and longitudinal fitting slopes of the subsequent trajectory at this scale, respectively.

[0050] In this embodiment, the fusion classification network takes 64-dimensional trajectory matching features and 64-dimensional physical embedding features as input, and concatenates the two types of features to form a 128-dimensional fusion feature. The fusion feature sequentially passes through a fully connected layer, an activation function, and a Dropout layer, and then through a binary classification output layer to obtain two output values ​​corresponding to discontinuous and continuous trajectories, respectively. Utilizing… The function converts the output value into category probabilities and uses the probability corresponding to the trajectory continuation category as the association confidence.

[0051] Furthermore, the trajectory encoder undergoes self-supervised pre-training through a trajectory encoding and denoising reconstruction network. During the pre-training phase, the original trajectory segments are diffused and denoised. The historical context features extracted by the trajectory encoder and the denoised trajectory features are input into the denoising decoder to reconstruct the trajectory data before denoising. The network parameters are updated based on the reconstruction error between the reconstructed trajectory and the original trajectory. After pre-training, the trajectory encoder is retained for downstream training and online inference, while the denoising decoder does not participate in trajectory continuation classification.

[0052] In this embodiment, the downstream training samples consist of preceding and subsequent trajectories. Trajectory pairs for the same target are labeled as positive samples, while trajectory pairs for different targets are labeled as negative samples. The online point-by-point growth process is simulated by truncating different length prefixes to the subsequent trajectories. The maximum training input lengths for preceding and subsequent trajectories are 200 and 64 points, respectively. Trajectories shorter than the preset length are padded with zeros, and these zeros are excluded using a valid point mask. The model is trained using binary cross-entropy loss. During online inference, the subsequent trajectory uses the currently received dynamically growing trajectory prefix, and the trajectory continuation category probability output by the model is used as the current association confidence level.

[0053] Step 5: Dynamically update the trajectory association status. Step 4 outputs the best associated target once each time. and the corresponding credibility Regarding the dynamic growth trajectory Association status Perform maintenance and verification. For each dynamically growing trajectory, maintain a fixed-length historical associated state queue. This is used to record the results of the most recent online association reasoning. Each step 4 outputs a new association result. Then, push the result into Tail of queue; when the queue length exceeds the preset independent decision window length. When this happens, the earliest record at the head of the queue is automatically discarded, making Always keep up with the latest This inference result avoids the storage overhead caused by the infinite growth of the sequence. In this embodiment, the window length is determined independently. Take 16.

[0054] Furthermore, following the order of stable association conditions, independent target conditions, and termination of association judgment, the current dynamic growth trajectory is assessed for association. When the most recent continuous... All online association results point to the same non-empty historical target, and the reliability of each association is not lower than the stable association threshold. When the condition for stable association is met, the associated information is determined to be... Set the target identity number of the historical target and associate the status. Set to Confirmed. In this embodiment, the stable association window length is... Set the threshold to 8 for stable association. Take 0.5.

[0055] When the stable association condition is not met, and the most recent consecutive All online association results are empty targets or their association confidence is lower than the independent target threshold. When the independent target condition is met, the associated information is determined. Mark as a new target and associate status Set to Confirmed. In this embodiment, the independent target threshold... Set the value to 0.4. When neither the stable correlation condition nor the independent objective condition is met, maintain... Continue receiving trajectory points and repeat steps 2-5.

[0056] Furthermore, when the dynamic growth trajectory reaches the preset maximum inference points, a termination association judgment is executed. This is if the last online association result points to a non-empty historical target, and the corresponding association credibility is not lower than the termination association threshold. Then the associated information will be Set it to the target identification number of this historical target; otherwise, associate information. Mark as a new target; after completing the above judgment, associate the status. Set to Confirmed. Specifically, when the confidence level of a dynamically growing association remains between the independence target threshold and the stable association threshold for an extended period, failing to meet either the stable association condition or the independence target condition, a termination of association decision will eventually be triggered. In this embodiment, the termination of association threshold... Set the value to 0.5, and the maximum number of reasoning points to 32.

[0057] Understandably, to avoid associating multiple dynamically growing trajectories with overlapping time ranges to the same historical target, when a trajectory meets the stable association condition, its time range is compared with a dynamically growing trajectory that has already been associated with that historical target but whose reception status is still "accepting". If the time ranges do not overlap, association with the same historical target is allowed; if the time ranges overlap, the association result of the first confirmed dynamically growing trajectory is retained, and that historical target is excluded from the subsequent candidate set of the later confirmed trajectory. Online association inference is then immediately re-executed based on the currently received trajectory points. The result of the re-inference is then included as a valid association result. The original association results are discarded, and the association judgment continues to be performed in the order of stable association condition, independent target condition, and termination association judgment.

[0058] Furthermore, once the target identity is confirmed, it cannot be revoked. Subsequent trajectory points are only used to complete the current trajectory segment and generate complete trajectory features, and will not change the already locked association information. For trajectories marked as new targets, a formal target number is assigned in step 6.

[0059] Step 6: Historical Target Update. When the reception status of the dynamic growth trajectory is "reception completed" and the association status is "confirmed", trajectory coding features, physical motion information, and trajectory cache data are generated based on the complete trajectory point sequence of the dynamic growth trajectory. If the association information is the target identity number of an existing historical target, the corresponding historical target record is updated using the generated information. If the association information is a new target marker, a new target identity number is assigned to the dynamic growth trajectory, a new historical target record is established, and it is added to the historical target management library.

[0060] In this embodiment, when the dynamic growth trajectory Reception status When this occurs, the historical target update process is triggered. The completion of reception includes not receiving a new trajectory point for more than a preset timeout threshold, or receiving a trajectory end marker from the data source. If at this time... If it is still in the process of being confirmed, then follow the termination association judgment in step 5 to determine its association information, and then... Updated to confirmed.

[0061] Furthermore, the data to be encoded is determined based on the complete trajectory segment, and when the complete trajectory segment exceeds the maximum input length of the model, the 200 most recent valid trajectory points at its end are extracted and input into the trajectory encoder to obtain the updated trajectory encoding features. Based on the valid trajectory points at the end of the trajectory, the end time, end position, velocity, and heading angle are calculated to obtain the updated physical motion information. Extract a preset number of trajectory points from the end of the complete trajectory to obtain the updated trajectory cache data. In this embodiment, the preset quantity is 200. When associated information... When pointing to an existing target in the historical target management database, the corresponding historical target record is updated using the generated trajectory encoding features, physical motion information, and trajectory cache data, while the target identification number of the target remains unchanged. When associated information When a target is marked as a new target, a new target identity number is assigned to it via a global target number generator. And build a new historical target record. It is then added to the historical target management database, and the newly assigned target identification number is written into the associated information. After the historical goals are updated, the dynamic growth trajectory will be displayed. Removed from the dynamically growing trajectory set. The updated historical target record serves as a historical reference for the subsequent real-time trajectory continuation of that target; when a new dynamically growing trajectory is formed, the target identity is still associated based on the historical target management library, so that subsequent trajectory segments belonging to the same target can continue to be connected to the corresponding target identity, thereby maintaining the continuity of target identity and trajectory information and realizing continuous tracking of the target.

[0062] To verify the effectiveness of the online trajectory association method, this embodiment conducts online and offline performance comparison experiments, as well as candidate screening, continuous confirmation, and historical feature delayed update ablation experiments in classic scenarios.

[0063] 1. Online vs. Offline Performance Comparison Experiment This experiment was conducted to verify the performance of the online point-to-point association method in terms of accurate identity association and early confirmation capabilities. Evaluation metrics included Precision, Recall, F1 score, false confirmation rate, average number of confirmed points, and average confirmation time. The false confirmation rate represents the proportion of trajectories whose target identity was incorrectly identified out of all test trajectories; the average number of confirmed points represents the average number of valid points received by a dynamically growing trajectory from the first valid point to the moment of first identity confirmation; and the average confirmation time represents the average time difference from the moment of the first valid trajectory point to the moment of first identity confirmation.

[0064] Table 1 shows the comparison results of offline 32-point association and online point-by-point association.

[0065] As shown in Table 1, the accuracy of the online method is lower than that of the offline method, but the F1 score and false confirmation rate are better, indicating that the present invention has good target identity association ability; the fewer confirmation points and shorter confirmation time indicate that the present invention can achieve effective early confirmation before the trajectory is received.

[0066] 2. Candidate screening ablation experiment This experiment was conducted to verify the effect of candidate target selection on association accuracy and online processing efficiency. Evaluation metrics included false confirmation rate, average number of confirmed cases, and average processing time per case.

[0067] Table 2 shows the comparison results of the two methods: retaining candidate screening and removing candidate screening.

[0068] Experimental results show that, as shown in Table 2, candidate screening can reduce the number of candidate trajectory pairs that the model needs to process and significantly reduce the average single-point processing time, which is beneficial to improving the efficiency of online correlation processing.

[0069] 3. Continuous confirmation ablation experiments This experiment was conducted to verify the impact of a continuous confirmation mechanism on confirmation speed and association stability. Evaluation metrics included average number of confirmations, F1 score, and false confirmation rate. The single-confirmation method refers to locking the associated identity upon first meeting a threshold.

[0070] Table 3 shows the comparison results of the effects of continuous confirmation and single confirmation methods.

[0071] Experimental results, as shown in Table 3, indicate that continuous confirmation can reduce erroneous associations caused by random reliability fluctuations with a certain confirmation wait time, thus providing support for reliable confirmation of online trajectories.

[0072] 4. Historical Feature Delayed Update Ablation Experiment This experiment was conducted to verify the effect of updating historical target features after trajectory reception on association stability. Evaluation metrics included false confirmation rate, associated target switching rate, and average single-point processing time. The associated target switching rate represents the proportion of times the best associated target in adjacent online association results changes before trajectory identity confirmation.

[0073] Table 4 shows the comparison results between two methods: delayed update after trajectory completion and immediate update after confirmation.

[0074] Experimental results, as shown in Table 4, indicate that the final identity association performance of the two historical feature update methods is similar. However, delaying the update after trajectory completion significantly reduces the switching degree of associated targets before identity confirmation. This suggests that maintaining historical target features unchanged during the dynamic growth trajectory reception process and updating them uniformly after the complete trajectory is formed helps reduce the interference of incomplete trajectory features on other online association processes, thereby improving the stability and efficiency of the association process.

[0075] The online trajectory association method described in this invention has good identity association and early confirmation capabilities. The candidate screening mechanism can improve online processing efficiency, and the continuous confirmation and historical feature delayed update mechanism can enhance the reliability and stability of the association process.

[0076] Example 2 This embodiment provides an online trajectory continuity system for real-time data streams, such as... Figure 2 As shown, it includes: The historical target management module is used to establish a historical target management library, assign a unique target number to each historical target, and save the trajectory coding features, trajectory physical motion information, and trajectory cache data corresponding to the most recent complete continuous trajectory segment of each historical target; before the dynamic growth trajectory reception is completed, the trajectory coding features, trajectory physical motion information, and trajectory cache data of each historical target in the historical target management library remain unchanged; The dynamic growth trajectory management module is used to create or update the corresponding dynamic growth trajectory based on the real-time trajectory identifier carried by the trajectory point each time a new valid trajectory point is received, and to continue receiving subsequent trajectory points after identity verification until the trajectory reception is completed. The candidate target dynamic filtering module is used to re-filter candidate historical targets whenever the dynamic growth trajectory is updated due to the addition of new valid trajectory points, before the identity of the dynamic growth trajectory has been confirmed. The online trajectory association reasoning module is used to perform online association reasoning once for each candidate historical target and the dynamically growing trajectory whenever the candidate target set is updated, and output the best associated target and its association confidence based on the association confidence. The trajectory association status dynamic update module is used to add each online association result to the association result sequence and perform an association judgment once. Before the trajectory is received, the target identity is confirmed based on multiple consecutive online association results, and the association information is locked after the target identity is confirmed. The historical target update module is used to keep the characteristics of historical targets in the historical target management library unchanged before the dynamic growth trajectory is received, and to update the corresponding historical target record or create a new historical target record based on the complete trajectory after the dynamic growth trajectory is received. The results output module is used to output the target number, association status, and association confidence level corresponding to the dynamic growth trajectory.

[0077] Furthermore, the online trajectory association reasoning module is equipped with a trajectory continuation neural network model, which includes a trajectory feature extraction network, a physical motion feature extraction network, and a fusion classification network. The trajectory feature extraction network includes a self-supervised pre-trained trajectory encoder, which generates current trajectory encoding features based on the trajectory feature tensor composed of lateral position, longitudinal position, lateral position change, and longitudinal position change, effective point mask, and actual trajectory length. The current trajectory encoding features are then combined with the historical trajectory encoding features of candidate historical targets read from the historical target management database, as well as the absolute difference features and product features of the two to obtain trajectory matching features. The physical motion feature extraction network is used to extract physical motion features under a unified local relative coordinate system based on the trajectory cache data of candidate historical targets, historical physical motion information, and the position coordinate sequence and sampling time of the current dynamically growing trajectory, and to generate physical embedding features. The fusion classification network is used to concatenate the trajectory matching features and the physical embedding features to form fusion features. After passing through a binary classification output layer and a Softmax function, the probability of the trajectory discontinuity category and the trajectory continuity category is obtained, and the probability of the trajectory continuity category is used as the association confidence.

[0078] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An online trajectory continuation method for real-time data streams, characterized in that, Includes the following steps: Step 1: Establish a historical target management database, assign a unique target identification number to each historical target, and save the historical trajectory coding features, historical physical motion information, and trajectory cache data generated from the most recently received complete trajectory segment of each historical target. Step 2: Receive the real-time input trajectory data stream. For each new trajectory point received, construct and update the dynamically growing trajectory based on the real-time trajectory identifier carried by the trajectory point. Step 3: When the target identity of the dynamically growing trajectory has not yet been confirmed, whenever a new trajectory point is added and the trajectory update is completed, based on the updated current physical motion information and the historical physical motion information of each historical target in the historical target management database, historical targets that meet the preset candidate conditions are selected to form a candidate target set. Step 4: When the target identity of the dynamically growing trajectory has not yet been confirmed, whenever a new trajectory point is added and the candidate target set is updated, online association reasoning is performed on the dynamically growing trajectory and each candidate historical target in the candidate target set, and the online association results are output. Step 5: Add each online association result to the association result sequence of the corresponding dynamic growth trajectory in the order of generation; perform identity judgment based on the continuous results in the association result sequence; after determining the target identity based on the judgment result, lock the association information and continue to receive subsequent trajectory points until the trajectory reception is completed; Step 6: When the receiving status of the dynamic growth trajectory is "received completed" and the association status is "confirmed", generate trajectory coding features, physical motion information and trajectory cache data based on the complete trajectory point sequence of the dynamic growth trajectory, and update the corresponding historical target record according to the confirmed association information to establish a new historical target record.

2. The online trajectory continuation method for real-time data streams according to claim 1, characterized in that, In step 2, the method for constructing and updating the dynamic growth trajectory is as follows: find the corresponding dynamic growth trajectory; if there is no corresponding trajectory, create a new dynamic growth trajectory and take the trajectory point as the first trajectory point; if there is already a corresponding trajectory, add the trajectory point to the trajectory point sequence; after each addition of a new trajectory point, update the physical motion information of the current trajectory synchronously according to the updated trajectory point sequence; each dynamic growth trajectory carries at least a real-time trajectory identifier, a trajectory point sequence, current physical motion information, an association status indicating whether the target identity has been confirmed, association information, and a trajectory reception status indicating whether the trajectory reception is complete.

3. The online trajectory continuation method for real-time data streams according to claim 1, characterized in that, In step 3, the preset candidate conditions include time interval constraints, spatial distance constraints, and motion consistency constraints; wherein: The time interval constraint is: the time interval between the start time of the dynamic growth trajectory and the end time of the most recent complete trajectory segment of the historical target does not exceed a preset time threshold. The spatial distance constraint is that the distance between the starting position of the dynamic growth trajectory and the ending position of the most recent complete trajectory segment of the historical target does not exceed a preset distance threshold. The motion consistency constraint is as follows: when the corresponding motion parameters are available, the difference between the velocity of the dynamic growth trajectory and the velocity of the historical target does not exceed a preset velocity difference threshold, and the difference between the heading angle of the dynamic growth trajectory and the heading angle of the historical target does not exceed a preset heading difference threshold; when the corresponding motion parameters are unavailable, the corresponding motion consistency constraint is not executed. Historical targets that simultaneously satisfy the time interval constraint, the spatial distance constraint, and the motion consistency constraint are added to the candidate target set.

4. The online trajectory continuation method for real-time data streams according to claim 1, characterized in that, In step 4, the method for online association reasoning is as follows: The current trajectory encoding features are generated based on the updated trajectory point sequence. The historical trajectory encoding features of candidate historical targets are read from the historical target management library, and the physical motion difference features are calculated based on the current physical motion information and the historical physical motion information of candidate historical targets. The current trajectory encoding features, the historical trajectory encoding features, and the physical motion difference features are fused and input into the trajectory continuation neural network model to obtain the predicted probability that the current dynamically growing trajectory and the candidate historical target belong to the same target, and this predicted probability is used as the association confidence. The candidate historical target with the highest association confidence is selected as the best association target for this time, and the online association result for this time is output.

5. The online trajectory continuation method for real-time data streams according to claim 4, characterized in that, In step 4, the trajectory continuation neural network model includes a trajectory feature extraction network, a physical motion feature extraction network, and a fusion classification network; The trajectory feature extraction network includes a self-supervised pre-trained trajectory encoder, which generates current trajectory encoding features based on the trajectory feature tensor composed of lateral position, longitudinal position, lateral position change, and longitudinal position change, effective point mask, and actual trajectory length. The current trajectory encoding features are then combined with the historical trajectory encoding features of candidate historical targets read from the historical target management database, as well as the absolute difference features and product features of the two to obtain trajectory matching features. The physical motion feature extraction network is used to extract physical motion features under a unified local relative coordinate system based on the trajectory cache data of candidate historical targets, historical physical motion information, and the position coordinate sequence and sampling time of the current dynamically growing trajectory, and to generate physical embedding features. The fusion classification network is used to concatenate the trajectory matching features with the physical embedding features to form fusion features. After passing through a binary classification output layer and a Softmax function, the class probabilities corresponding to the trajectory discontinuity category and the trajectory continuity category are obtained, and the probability corresponding to the trajectory continuity category is used as the association confidence.

6. The online trajectory continuation method for real-time data streams according to claim 1, characterized in that, In step 5, the method for determining identity based on consecutive results in the association result sequence is as follows: when multiple consecutive association results meet the preset stable association conditions, the association information of the current dynamic growth trajectory is set as the target identity number of the corresponding historical target, and the association status is updated to confirmed; when the preset independent target condition is met, the association information of the current dynamic growth trajectory is marked as a new target, and the association status is updated to confirmed; when none of the above conditions are met, the association status is kept as unconfirmed, and steps 2 to 5 are continued after a new trajectory point is added; after the target identity is determined, the association information is locked, the candidate target screening and online association reasoning for identity selection are stopped for the dynamic growth trajectory, and subsequent trajectory points are received until the trajectory reception is completed.

7. The online trajectory continuation method for real-time data streams according to claim 6, characterized in that, In step 5, the stable association condition includes the target number consistency condition and the association credibility stability condition; a judgment window length and an association credibility threshold are set. If the best associated target number remains consistent in the most recent consecutive association inference results of the judgment window length, and the association credibility is higher than the association credibility threshold each time within the window, then the stable association condition is determined to be satisfied.

8. The online trajectory continuation method for real-time data streams according to claim 6, characterized in that, In step 5, the independent target condition is: set the judgment window length and the independent judgment threshold. If the association credibility of the most recent consecutive association reasoning results of the judgment window length is lower than the independent judgment threshold, then the current trajectory is determined not to belong to any existing historical target, thus satisfying the independent target condition. The zero association credibility result generated when the candidate target set is empty in step 4 participates in the continuous judgment in the same way.

9. The online trajectory continuation method for real-time data streams according to claim 6, characterized in that, In step 5, when the dynamic growth trajectory is still in an unconfirmed state, and the number of trajectory points reaches the preset maximum inference points or the trajectory reception status changes to reception completed, a termination association judgment is executed. The termination association judgment is as follows: read the last online association result, and if the best candidate historical target is not empty and the corresponding association credibility is not lower than the preset termination association threshold, set the association information to the target identity number corresponding to the historical target; otherwise, mark the association information as a new target and update the association status to confirmed.

10. The online trajectory continuation method for real-time data streams according to claim 6, characterized in that, In step 5, when a trajectory meets the stable association condition, the time range of the dynamically growing trajectory is compared with the time range of other dynamically growing trajectories that have been associated with the same historical target. If the time ranges do not overlap, association with the same historical target is allowed. If the time ranges overlap, the association result of the previously confirmed trajectory is retained, and the historical target is excluded from the candidate set of the subsequently confirmed trajectory. Online association reasoning is re-executed based on the currently received trajectory points. The result of the re-reasoning is added to the association result sequence as a valid association result, the original association result is discarded, and the association judgment continues to be performed in the order of stable association condition, independent target condition, and termination of association judgment.

11. The online trajectory continuation method for real-time data streams according to claim 1, characterized in that, In step 1, before the dynamic growth trajectory reception is completed, the historical trajectory encoding features, historical physical motion information and trajectory cache data of each historical target in the historical target management library are kept unchanged.

12. The online trajectory continuation method for real-time data streams according to claim 1, characterized in that, In step 2, when a trajectory end marker is received or no new trajectory point is received within a preset time, the trajectory receiving status of the corresponding dynamically growing trajectory is updated to reception completed.

13. The online trajectory continuation method for real-time data streams according to claim 1, characterized in that, In step 6, if the associated information is the target identity number of an existing historical target, the corresponding historical target record is updated using the generated information; if the associated information is a new target marker, a new target identity number is assigned to the dynamic growth trajectory, a new historical target record is established and added to the historical target management database.

14. An online trajectory continuity system for real-time data streams, characterized in that, include: The historical target management module is used to establish a historical target management library, assign a unique target number to each historical target, and save the trajectory coding features, trajectory physical motion information, and trajectory cache data corresponding to the most recent complete continuous trajectory segment of each historical target; before the dynamic growth trajectory reception is completed, the trajectory coding features, trajectory physical motion information, and trajectory cache data of each historical target in the historical target management library remain unchanged; The dynamic growth trajectory management module is used to create or update the corresponding dynamic growth trajectory based on the real-time trajectory identifier carried by the trajectory point each time a new valid trajectory point is received, and to continue receiving subsequent trajectory points after identity verification until the trajectory reception is completed. The candidate target dynamic filtering module is used to re-filter candidate historical targets whenever the dynamic growth trajectory is updated due to the addition of new valid trajectory points, before the identity of the dynamic growth trajectory has been confirmed. The online trajectory association reasoning module is used to perform online association reasoning once for each candidate historical target and the dynamically growing trajectory whenever the candidate target set is updated, and output the best associated target and its association confidence based on the association confidence. The trajectory association status dynamic update module is used to add each online association result to the association result sequence and perform an association judgment once. Before the trajectory is received, the target identity is confirmed based on multiple consecutive online association results, and the association information is locked after the target identity is confirmed. The historical target update module is used to keep the characteristics of historical targets in the historical target management library unchanged before the dynamic growth trajectory is received, and to update the corresponding historical target record or create a new historical target record based on the complete trajectory after the dynamic growth trajectory is received. The results output module is used to output the target number, association status, and association confidence level corresponding to the dynamic growth trajectory.

15. The online trajectory continuity system for real-time data streams according to claim 14, characterized in that, The online trajectory association reasoning module is equipped with a trajectory continuation neural network model, which includes a trajectory feature extraction network, a physical motion feature extraction network, and a fusion classification network. The trajectory feature extraction network includes a self-supervised pre-trained trajectory encoder, which generates current trajectory encoding features based on the trajectory feature tensor composed of lateral position, longitudinal position, lateral position change, and longitudinal position change, effective point mask, and actual trajectory length. The current trajectory encoding features are then combined with the historical trajectory encoding features of candidate historical targets read from the historical target management database, as well as the absolute difference features and product features of the two to obtain trajectory matching features. The physical motion feature extraction network is used to extract physical motion features under a unified local relative coordinate system based on the trajectory cache data of candidate historical targets, historical physical motion information, and the position coordinate sequence and sampling time of the current dynamically growing trajectory, and to generate physical embedding features. The fusion classification network is used to concatenate the trajectory matching features and the physical embedding features to form fusion features. After passing through a binary classification output layer and a Softmax function, the probability of the trajectory discontinuity category and the trajectory continuity category is obtained, and the probability of the trajectory continuity category is used as the association confidence.

Images