An unmanned aerial vehicle track correlation method based on multi-source remote identification signals

Abstract

The application discloses a kind of unmanned plane track association methods based on multi-source remote identification signal, it is related to low-altitude flight supervision and target tracking technical field.The specific implementation scheme is: obtaining the multi-source observation data in target area and executing pre-processing to obtain candidate observation data;Dynamic association threshold is generated based on signal quality parameter and time parameter to screen candidate matching data;The multi-dimensional association cost of fusion space, motion, time and identity joint consistency between candidate matching data and track prediction state is calculated;Multi-dimensional association cost is mapped to single frame confidence and time sequence recursion is executed to obtain current cumulative confidence;In response to current cumulative confidence satisfying preset confirmation condition, weighted fusion is carried out on prediction state and observation state based on single frame confidence to update track state, by the application, the accuracy and continuity of unmanned plane track association under complex occlusion environment are effectively improved.

Description

Technical Field

[0001] This application relates to the field of low-altitude flight monitoring and target tracking technology, specifically to a method for associating UAV tracks based on multi-source remote identification signals. Background Technology

[0002] With the widespread application of drones in logistics delivery, power line inspection, emergency rescue, and urban management, the demand for low-altitude airspace safety supervision is increasing. In order to effectively monitor drone flight activities, countries have successively implemented remote identification systems, requiring drones to continuously broadcast remote identification messages containing information such as identity, location, altitude, speed, and heading during flight. Supervision platforms typically collect these broadcast signals by deploying multiple ground receiving stations within urban areas and construct continuous and stable drone flight paths based on them to support flight monitoring, behavior analysis, and violation warnings.

[0003] Track association is a core component of UAV track construction, its task being to correctly assign discrete observation data from different receiving stations at different times to the corresponding UAV tracks. Under ideal conditions—that is, when receiving station clocks are synchronized, signal propagation is unobstructed, and UAV identification is broadcast stably—track association can be achieved through simple nearest neighbor matching or a fixed distance threshold. However, in densely populated urban areas with tall buildings, actual observation conditions are far from ideal, and track association faces multiple technical challenges.

[0004] First, there is the issue of asynchronous timing in multi-source data. There are fixed system deviations and random transmission jitter between the local clocks of different receiving stations and the drone's broadcast clock, resulting in inconsistent timestamps reported by the same drone at different stations. Without unified correction, observation data from different stations cannot be effectively compared on the same timeline, leading to the same target being incorrectly split into multiple targets.

[0005] Second, data quality issues caused by signal obstruction and multipath effects. Blockage from tall buildings can lead to message loss, delays, and duplicate reception, while multipath reflections can cause drastic fluctuations in signal strength and positional information drift. Within a short time window at the same receiving station, multiple redundant messages from the same UAV may be received due to protocol retransmissions or multipath effects. Without deduplication, this will result in an abnormally high track update frequency and even oscillations in state estimation.

[0006] Third, there is the issue of abnormal changes in identification identifiers. Drones may experience temporary changes in their remote identification number during flight due to unstable communication links, switching of privacy protection policies, or device restarts. Traditional flight path association methods rely heavily on identification identifier matching; once the identifier changes, the same drone may be incorrectly identified as a new target, creating a new flight path and fragmenting what was originally a continuous flight trajectory into multiple segments.

[0007] Fourth, the problem of distinguishing adjacent targets in dense scenes. In urban low-altitude areas, multiple drones may fly simultaneously in the same airspace, and their motion parameters such as altitude, speed, and heading may be very similar. Relying solely on nearest neighbor or fixed distance thresholds for matching can easily lead to incorrect association of observation data from adjacent drones to the same track, causing track confusion. Fifth, the unreliability of single-frame observations. Due to the influence of instantaneous noise, multipath reflection, or brief occlusion, the observation data of a single frame may deviate significantly from the true value. If correlations are confirmed based on a single observation, the probability of misjudgment will increase significantly, leading to the establishment of false tracks or the contamination of true tracks.

[0008] To address the aforementioned issues, existing technologies primarily employ the following methods for track association: Nearest neighbor methods based on fixed distance thresholds, which are simple to implement but cannot adapt to dynamically changing observation conditions. When signal obstruction leads to increased noise or message loss, fixed thresholds can easily cause false associations or track breaks; Kalman filter-based tracking methods, which typically assume that noise follows a Gaussian distribution with known or slowly changing statistical characteristics. However, in urban low-altitude surveillance scenarios, the error distribution of UAV remote identification signals often exhibits heavy-tailed, asymmetric, or even multimodal characteristics, and fluctuates drastically with the degree of obstruction. This results in excessive corrections from adaptive gains based on model covariance under strong interference or insufficient corrections under weak signals; and methods based on multi-hypothesis tracking or joint probability data association. While these methods can theoretically handle complex scenarios, their computational complexity increases exponentially with the number of targets, making it difficult to meet real-time requirements and incurring high deployment costs.

[0009] In summary, existing flight path association methods, when faced with multiple challenges in urban environments such as multi-source asynchronous data, signal quality fluctuations, abrupt changes in identification markers, and similar parameters of adjacent targets, either lack robustness leading to frequent false associations and track breaks, or are too complex to be practically implemented. Therefore, there is an urgent need for a UAV flight path association method that can maintain high robustness and continuity under conditions of multi-source asynchronous data, signal fluctuations, and identification anomalies. Summary of the Invention

[0010] In view of this, the purpose of this application is to provide a UAV track association method based on multi-source remote identification signals, which can realize accurate, continuous association and stable update of UAV tracks, and effectively solve the problems of asynchronous multi-source observation data, signal quality fluctuations and abnormal identification in complex obstruction environments, which lead to erroneous track association and track breakage.

[0011] According to a first aspect of this application, a method for associating unmanned aerial vehicle (UAV) tracks based on multi-source remote identification signals is provided, comprising: Acquire multi-source observation data within the target area and perform preprocessing on the multi-source observation data to obtain candidate observation data; Based on the signal quality parameters and time parameters corresponding to the candidate observation data, a dynamic correlation threshold is generated, and the candidate matching data that meets the spatial difference requirements is selected from the candidate observation data using the dynamic correlation threshold. Calculate the multidimensional association cost between candidate matching data and the predicted state of the corresponding track. The multidimensional association cost integrates spatial differences, motion differences, temporal differences, and joint consistency of identity and motion. The multidimensional association cost is mapped to a single-frame confidence score, and the single-frame confidence score is recursively calculated based on the historical cumulative confidence score to obtain the current cumulative confidence score. In response to the current cumulative confidence meeting the preset confirmation conditions, the candidate matching data is confirmed to belong to the corresponding track, and the predicted state of the corresponding track and the observed state of the candidate matching data are weighted and fused based on the single frame confidence, and the state of the corresponding track is updated.

[0012] In some embodiments, a dynamic correlation threshold is generated based on the signal quality parameters and time parameters corresponding to the candidate observation data, including: The time difference between candidate observation data and the predicted state of the corresponding track is obtained, and a time difference compensation term is generated based on the time difference. The time difference compensation term is positively correlated with the time difference. A signal quality compensation term is generated based on the signal quality parameters corresponding to the candidate observation data. The signal quality compensation term is negatively correlated with the signal quality parameters. The historical state uncertainty of the corresponding track is obtained, and an uncertainty compensation term is generated based on the historical state uncertainty. The uncertainty compensation term is positively correlated with the historical state uncertainty. The preset basic threshold, time difference compensation term, signal quality compensation term, and uncertainty compensation term are superimposed to obtain the dynamic correlation threshold.

[0013] In some embodiments, when calculating the multidimensional association cost between candidate matching data and the predicted state of the corresponding track, the evaluation logic for joint identity and motion consistency includes: When the candidate matching data is completely consistent with the identity identifier of the corresponding track, the first identity value is assigned. When the candidate matching data is inconsistent with the identity of the corresponding track, and the spatial and temporal differences between the two are within the preset tolerance range, a second identity value is assigned, and the value of the second identity value is greater than that of the first identity value. When the candidate matching data is inconsistent with the identity of the corresponding track, and the spatial or temporal difference between the two exceeds the preset tolerance range, a third identity value is assigned, and the value of the third identity value is greater than the value of the second identity value. The value of the first identity, the value of the second identity, or the value of the third identity are used as identity consistency costs to participate in the calculation of multidimensional association costs.

[0014] In some embodiments, calculating the multidimensional association cost between candidate matching data and the predicted state of the corresponding track further includes: Calculate the difference in planar distance in the planar position, the difference in height in the vertical direction, the difference in velocity in the motion speed, and the difference in time in the observation time between the candidate matching data and the predicted state of the corresponding track. Extract the first heading angle of the candidate matching data and the second heading angle of the predicted state of the corresponding track, calculate the absolute difference between the first heading angle and the second heading angle, and compare the absolute difference with the result of subtracting the absolute difference from the preset period threshold, and take the smaller value of the two as the heading difference; The plane distance difference, altitude difference, speed difference, heading difference, and time difference are normalized using the corresponding normalization constants to obtain the normalized difference values ​​for each dimension. The multidimensional association cost is obtained by weighted summation of the normalized difference values ​​of each dimension and the identity consistency cost.

[0015] In some embodiments, the multidimensional association cost is mapped to a single-frame confidence score, and a temporal recursion is performed on the single-frame confidence score based on the historical cumulative confidence score to obtain the current cumulative confidence score, including: The multidimensional association cost is mapped to a single-frame confidence level between zero and one by using an exponential decay function. The multidimensional association cost and the single-frame confidence level are negatively correlated. Obtain the historical cumulative confidence level of the corresponding track at the previous moment; The historical cumulative confidence and single-frame confidence are weighted and fused using a preset forgetting factor to obtain the current cumulative confidence. The preset forgetting factor is used to balance historical continuity and current frame changes.

[0016] In some embodiments, in response to the current accumulated confidence level meeting a preset confirmation condition, confirming that the candidate matching data belongs to the corresponding track includes: Set an acceptance threshold and an rejection threshold. If the acceptance threshold is greater than the rejection threshold, a status buffer will be formed. When the current cumulative confidence level is greater than or equal to the confirmation threshold, the candidate matching data is confirmed to belong to the corresponding track, and the weighted fusion is triggered to update the status of the corresponding track. When the current cumulative confidence level is less than or equal to the elimination threshold, candidate matching data is eliminated or the association with the corresponding track is broken; When the current cumulative confidence level is between the elimination threshold and the confirmation threshold, the pending state of the candidate matching data is maintained, and the status of the corresponding track is not updated for the time being.

[0017] In some embodiments, the predicted state of the corresponding track and the observed state of the candidate matching data are weighted and fused based on the single-frame confidence to update the state of the corresponding track, including: The confidence level of a single frame is determined as the observation weight, and the difference between the observation weight and the prediction weight is determined as the prediction weight. The observation states of the candidate matching data are weighted using observation weights to obtain the first weighted state; The predicted states of the corresponding trajectories are weighted using the prediction weights to obtain the second weighted state; The first weighted state is added to the second weighted state to obtain the fused track state, which makes the dependence ratio of the observed state positively correlated with the single-frame confidence, and the dependence ratio of the predicted state negatively correlated with the single-frame confidence.

[0018] In some embodiments, multi-source observation data undergoes preprocessing to obtain candidate observation data, including: Within the preset calibration window, the average deviation between the reception time and the broadcast time of the same data source is calculated. The average deviation is used to correct the reception time of subsequent multi-source observation data, and the time-synchronized observation data is output. Within the preset duplicate detection window, multiple duplicate data entries from the same data source and with the same identity are filtered out from the time-synchronized observation data. Compare the signal quality parameters of multiple duplicate data points, retain only the duplicate data point with the highest signal quality parameter as the valid observation data, delete the remaining duplicate data, and use the valid observation data as the candidate observation data.

[0019] In some embodiments, after updating the state of the corresponding track, the method further includes: Based on the uniform motion model, the predicted plane position and predicted altitude at the next moment are calculated using the current speed, heading angle and time interval of the corresponding track. Extract the historical heading angles of the most recent consecutive frames of the corresponding track, and calculate the sum of the sine and cosine values ​​of each historical heading angle respectively; The smoothed heading angle is obtained by calculating the sum of the sine and cosine values ​​using a two-dimensional arctangent function, thus eliminating periodic angle jumps.

[0020] In some embodiments, when no candidate matching data that meets the dynamic association threshold and preset confirmation conditions is found within the current time period, the corresponding track is marked as lost and the number of consecutive lost frames is incremented by one. Determine if the number of consecutive lost connections exceeds the preset loss limit; If the number of consecutive lost frames does not exceed the preset loss limit, the prediction status of the corresponding track will be maintained and it will continue to participate in subsequent correlations. If the number of consecutive lost connections exceeds the preset loss limit, the corresponding flight path will be terminated and associated resources will be released.

[0021] One embodiment of the above application has the following advantages or beneficial effects: This application constructs a multi-dimensional dynamic correlation evaluation model that integrates signal quality, time deviation, and identity-motion joint consistency. This model enables the correlation threshold to be adaptively adjusted according to real-time changes in the observation environment, effectively overcoming the problems of false correlation and missed correlation that are easily caused by fixed thresholds under signal fluctuation or time asynchrony conditions, and improving the matching robustness in complex occlusion environments.

[0022] This application introduces a multi-frame confidence accumulation mechanism based on temporal recursion and a dual-threshold buffer judgment strategy, which transforms the single-frame association result into a temporal confidence metric, avoids single-frame misjudgment caused by instantaneous noise or multipath interference, significantly reduces the probability of track breakage and false association, and ensures the continuity and stability of the track.

[0023] This application adopts an adaptive weighted fusion update strategy based on single-frame confidence, which enables the track state update process to dynamically adjust the dependence ratio of the observed and predicted values ​​according to the confidence level of the current observation. While preserving the historical motion trend, it reasonably absorbs new observation information, effectively suppresses track jumps, and improves the accuracy of state estimation.

[0024] This application adopts a joint hierarchical evaluation mechanism based on identity identification and motion continuity. It assigns the optimal matching weight when the identity identification is consistent, retains a moderate correlation possibility when the identity identification is abnormal but the motion state is continuous, and imposes a high penalty when neither identity nor motion matches. This balances the discriminative power of identity information with the constraint of motion continuity, and solves the problem of incorrect track splitting caused by temporary changes in identity identification.

[0025] Other effects of the above-mentioned alternative methods will be described below in conjunction with specific embodiments. Attached Figure Description

[0026] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein: Figure 1This is a flowchart illustrating a method for associating UAV tracks based on multi-source remote identification signals, provided in an embodiment of this application. Figure 2 This is a schematic diagram of a tracking trajectory association provided in an embodiment of this application; Figure 3 This is a schematic diagram of the costs in various dimensions in a typical scenario of temporary identity change provided in the embodiments of this application; Figure 4 This is a schematic diagram of the confidence level change curve in a typical scenario of temporary identity change provided in the embodiments of this application; Figure 5 This is a schematic diagram of typical confidence curves of a drone passing through an open area, a blocked area, and a recovery area, provided in an embodiment of this application. Detailed Implementation

[0027] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0028] To facilitate understanding of this application, the embodiments of this application will be briefly described below: The method described in this application requires the use of multiple preset parameters during operation. These parameters are not arbitrarily set, but rather determined through a combination of offline calibration and online verification, based on the physical characteristics of the UAV remote identification signal, the statistical patterns of typical flight scenarios, and the engineering requirements of system performance indicators. The parameters are explained below in the order they appear in the method flow.

[0029] The first category consists of parameters used in the data preprocessing stage. A preset calibration window is used to statistically analyze the time deviation of the receiving station. Its value needs to balance the statistical sample size and timeliness: if the window is too short, the sample size will be insufficient to eliminate random jitter; if the window is too long, it will be unable to track changes in clock drift. In urban low-altitude surveillance scenarios, the broadcast frequency of UAV remote identification messages is typically 1 Hz, i.e., once per second. Therefore, setting the preset calibration window to 10 seconds allows for the accumulation of approximately 10 time difference samples within the window, sufficient to filter out random transmission jitter and extract stable clock offsets through mean estimation. If the broadcast frequency is higher, the window can be appropriately shortened; if the clock drift is faster, the window should be shortened to improve tracking capability. A preset duplicate detection window is used to identify duplicate messages received by the same receiving station. Its value should cover the maximum time spread caused by communication protocol retransmissions and multipath reflections. In urban environments, message delays caused by multipath reflections are typically between tens and hundreds of milliseconds. Therefore, setting the preset duplicate detection window to 500 milliseconds covers the vast majority of duplicate messages without misjudging normal broadcasts from the same UAV at different times as duplicates. A preset available threshold is used to filter invalid messages with too low a signal-to-noise ratio. Its value depends on the receiver's sensitivity limit and decoding reliability requirements. It is usually set to the lowest signal-to-noise ratio threshold at which the receiver can correctly decode the message content, such as 0 dB or 3 dB. Messages below this threshold are directly rejected because the decoding error rate is too high.

[0030] The second category consists of parameters in the dynamic association threshold generation stage. The preset base threshold is the baseline component of the dynamic threshold, representing the maximum allowable spatial association range under ideal observation conditions (i.e., zero time difference, extremely high signal-to-noise ratio, and completely determined track status). The calibration method for this value is as follows: Collect a large amount of continuous observation data confirmed to belong to the same track in an open, unobstructed environment; statistically analyze the planar distance distribution between candidate observations and predicted states; and take the 95th percentile of this distribution as the preset base threshold. For example, when the UAV's flight speed is 10 to 20 meters per second and the message broadcast frequency is 1 Hz, the maximum reasonable displacement between two adjacent frames is approximately 20 meters. In this case, the preset base threshold can be set to 20 meters. The generation of the time difference compensation term involves a time difference weighting coefficient. The physical meaning of this coefficient is how many extra meters the association threshold needs to be widened to tolerate the possible displacement of the target during each additional second of time asynchrony. The calibration method is as follows: Using test data from known flight paths, statistically analyze the actual spatial deviation between the observed and predicted points under different time difference conditions. Then, fit the slope of the spatial deviation as a function of the time difference using linear regression. This slope is the time difference weighting coefficient. For example, when the typical cruising speed of the UAV is 15 meters per second, the time difference weighting coefficient can be calibrated to 15 meters per second, meaning that for every additional second of time difference, the threshold is widened by an additional 15 meters. The generation of the signal quality compensation term involves the signal-to-noise ratio (SNR) weighting coefficient. The physical meaning of this coefficient is that when the SNR decreases, the possible drift of the observed position increases, and the threshold needs to be widened accordingly. The calibration method is as follows: Under different SNR conditions, statistically analyze the deviation distribution between the position decoded by the receiver and the actual position, establish the mapping relationship between SNR and position deviation, and then determine the SNR weighting coefficient accordingly. For example, when the SNR decreases from 15 dB to 5 dB, the standard deviation of the position drift may increase from 2 meters to 10 meters. In this case, the SNR weighting coefficient should reflect this increase. The generation of the uncertainty compensation term involves an uncertainty weighting coefficient, which reflects the strength of the influence of the reliability of the track state estimation on the threshold. Track uncertainty is obtained by statistically analyzing the variance or standard deviation of the observation residuals (i.e., the difference between observed and predicted values) of the track over the most recent few frames. Large fluctuations in the residuals indicate unstable track state estimation, requiring more lenient matching conditions. The uncertainty weighting coefficient is calibrated by using a grid search to determine the uncertainty weighting value that minimizes the track break rate in test scenarios involving frequent maneuvers and obstructions, with track tracking continuity as the optimization objective. The zero-prevention minimum is an extremely small positive number used to prevent division-by-zero errors in the signal quality compensation term when the signal-to-noise ratio is zero. It is typically set to a value much smaller than the normal signal-to-noise ratio range, such as 0.001.

[0031] The third category consists of parameters used in the multidimensional correlation cost calculation stage. Normalization constants for each dimension are used to eliminate differences in the dimensions and orders of magnitude of different physical quantities, enabling fair comparison of heterogeneous features such as position, velocity, heading, and time on the same scale. The calibration methods for each normalization constant are as follows: The altitude normalization constant is taken as the maximum reasonable altitude difference of the drone within the monitored area. For example, the upper limit of urban low-altitude monitoring altitude is usually 120 meters, so the altitude normalization constant can be set to 120 meters or determined based on the actual upper limit of the monitored airspace altitude; the velocity normalization constant is taken as the maximum flight speed of the drone. For example, the maximum horizontal speed of a consumer-grade drone is usually 20 meters per second, so the velocity normalization constant can be set to 20 meters per second; the heading normalization constant is taken as the maximum possible difference in heading angle, i.e., 180 degrees, because the maximum value of the heading difference after minimum orbital difference processing will not exceed 180 degrees; the time normalization constant is taken as the maximum allowable time asynchrony, for example, determined based on the message broadcast frequency and the maximum allowable delay, usually set to 1 second. The weight coefficients for each dimension are used to adjust the relative importance of different dimensions in the total association cost, and the sum of all weight coefficients equals 1. The weight calibration method is as follows: First, on the labeled test dataset (i.e., data where the true attribution of each observation is known), with the optimization objective of minimizing the false association rate, the optimal combination in the weight space is searched using a grid search or Bayesian optimization algorithm; then, fine-tuning is performed according to the actual scenario requirements. For example, in densely populated urban areas with tall buildings, due to large positional drift but relatively stable heading, the position weight can be appropriately reduced and the heading weight increased; in open airspace, where positional information is more reliable, the position weight can be appropriately increased. A typical weight configuration is: planar position weight 0.30, altitude weight 0.15, velocity weight 0.15, heading weight 0.15, time weight 0.10, and identity consistency weight 0.15. The three tiered values ​​in the identity consistency cost also need to be calibrated: the first identity cost value (when identities are completely consistent) is set to 0, indicating no penalty is imposed; the second identity cost value (when identities are inconsistent but movement is continuous) is set to an intermediate value greater than 0 and less than 1, such as 0.3. The calibration principle for this value is to reflect the risk of identity anomalies (therefore it cannot be 0), but it cannot be so large that it excludes the same target with highly continuous movement (therefore it is much less than 1). Specifically, it can be determined by statistically analyzing the proportion of movement continuity that is established in identity jump events; the third identity cost value (when identities are inconsistent and movement is discontinuous) is set to 1, representing the maximum penalty, to effectively suppress the misassociation of other nearby UAVs with the current track. The preset tolerance range in the identity consistency determination includes spatial tolerance and temporal tolerance: the spatial tolerance takes the current value of the dynamic association threshold, that is, when the difference between the plane distance between the candidate observation and the track is less than or equal to the current dynamic threshold, it is considered spatially continuous; the temporal tolerance takes a time normalization constant, such as 1 second, that is, when the time difference between the candidate observation and the track is less than or equal to 1 second, it is considered temporally continuous.

[0032] The fourth category consists of parameters for the confidence accumulation and state determination stages. A preset forgetting factor is used to balance the weighting of historical accumulated confidence and current single-frame confidence, with a value range greater than 0 and less than 1. The forgetting factor is calibrated as follows: on test data containing typical occlusion events (such as a drone flying into and out of the shadow of a tall building), the optimal value is determined by iterating through different forgetting factor values, using a comprehensive index of track continuity and false association rate as the optimization objective. If the forgetting factor is too large (e.g., above 0.95), it relies too much on historical information, resulting in a delayed response when the target undergoes real maneuvers or a legitimate change in identity; if the forgetting factor is too small (e.g., below 0.5), it is too sensitive to the current frame and easily susceptible to false confirmations due to sudden interference. In densely built-up urban areas, occlusion events typically last for several seconds, so the forgetting factor should be set between 0.7 and 0.9, such as 0.85. At this value, observations within the most recent approximately 7 frames (i.e., approximately 7 seconds) are given higher weight, which can smooth out instantaneous noise and respond promptly when the target's behavior changes. The confirmation threshold is the limit for determining whether an association is formally established. Its calibration method is to statistically analyze the cumulative confidence distribution of correct and incorrect associations on labeled data, and select a value that achieves the optimal balance between the recall rate of correct associations and the false association rate. For example, setting the confirmation threshold to 0.8 means that the average confidence of multiple consecutive frames must reach above 80% before a formal association is confirmed, which typically requires at least 2 to 3 frames of high-quality matches. The rejection threshold is the limit for determining whether an association is invalid. Its calibration principle is to form a sufficiently wide buffer between the confirmation threshold and the rejection threshold to avoid frequent state switching when the cumulative confidence fluctuates near the critical value. For example, setting the rejection threshold to 0.3 results in a buffer width of 0.5 between the confirmation threshold and the rejection threshold, which is sufficient to absorb confidence jitter caused by short-term fluctuations in signal quality. The specific value of the buffer width can be determined by statistically analyzing the standard deviation of confidence fluctuations between consecutive frames, and is generally set to 2 to 3 times the standard deviation of fluctuations.

[0033] The fifth category is parameters for the loss of contact management phase. The preset loss of contact upper limit determines the maximum retention time of a track without observation updates. This parameter is calibrated by statistically analyzing the duration distribution of typical occlusion events in the target scene and taking the 90th percentile of this distribution as the preset loss of contact upper limit. For example, in densely populated urban areas with tall buildings, the duration of most building occlusion events does not exceed 5 seconds (corresponding to 5 frames at a 1 Hz broadcast frequency), therefore the preset loss of contact upper limit is set to 5 frames. If the setting is too small (e.g., 1 to 2 frames), it will not cover common occlusion durations, leading to frequent track drops; if the setting is too large (e.g., more than 20 frames), targets that have already flown away or landed will still occupy resources as ghost tracks, even interfering with the establishment of new targets. In open airspace or scenarios with less occlusion, this value can be appropriately reduced to improve resource utilization efficiency.

[0034] It is important to note that the calibration of the above parameters is not a one-time static process, but rather can be periodically calibrated and optimized based on actual operational data after deployment. For example, when the receiving station hardware is replaced or the deployment location is adjusted, the time deviation within the preset calibration window should be recalibrated; when the type or flight density of drones in the monitored area changes significantly, the weighting coefficients and threshold parameters should be re-optimized. Furthermore, there are coupling relationships between the parameters; adjusting one parameter may require adjusting other parameters in conjunction to maintain overall performance. For example, if the preset base threshold is increased, the dynamic threshold is relaxed overall. In this case, the confirmation threshold can be appropriately increased to maintain the strictness of the association and avoid an increase in the false association rate.

[0035] See Figure 1 This is a flowchart illustrating a method for UAV trajectory association based on multi-source remote identification signals provided in an embodiment of this application. Figure 1 The execution subject of the method shown can be a combination of software and / or hardware, specifically, it can be one or more of various types of terminals, hardware systems, cloud computing, etc.

[0036] Figure 1 The method for UAV track association based on multi-source remote identification signals, shown below, includes steps S101 to S105. By constructing a complete closed loop from data preprocessing to adaptive state update, it solves the problems of track misassociation and track breakage caused by asynchronous multi-source data, signal fluctuations, and identity anomalies in complex occlusion environments. The details are as follows: S101: Acquire multi-source observation data within the target area and perform preprocessing on the multi-source observation data to obtain candidate observation data.

[0037] Specifically, multi-source observation data refers to raw detection information about drones collected by multiple heterogeneous or homogeneous sensors deployed within a target area. For example, in urban low-altitude surveillance scenarios, this data may originate from fixed receiving stations deployed on rooftops, mobile receiving terminals mounted on inspection vehicles, or even other IoT devices with sensing capabilities. Because different sensors often have different clock references, communication protocols, and data formats, and the raw data frequently contains spurious signals due to multipath effects or outliers caused by equipment malfunctions, preprocessing is essential. Preprocessing functions mainly include, but are not limited to, time reference unification, duplicate data removal, and invalid anomaly filtering.

[0038] This process transforms the chaotic raw data from different sources and at different times into candidate observation data with consistent spatiotemporal references and relatively reliable quality, laying the data foundation for subsequent accurate correlation. If this step is omitted or improperly processed, subsequent correlation algorithms will face severe data pollution, leading to wasted computational resources or even erroneous correlation results.

[0039] S102: Based on the signal quality parameters and time parameters corresponding to the candidate observation data, generate a dynamic correlation threshold, and use the dynamic correlation threshold to filter out candidate matching data that meet the spatial difference requirements from the candidate observation data.

[0040] The core of this step lies in abandoning the fixed distance or time thresholds of traditional methods and instead adopting a dynamic filtering mechanism that can adapt to the environment. Specifically, the signal quality parameter reflects the reliability of the current observation data, while the time parameter characterizes the time span between the observation data and the predicted trajectory. When the signal quality is poor or the time delay is large, it means that the uncertainty of the observation increases. At this time, the correlation threshold will be automatically relaxed to avoid missing the real target due to overly strict filtering. Conversely, when the signal is clear and the time synchronization is good, the threshold will be tightened to improve the anti-interference capability.

[0041] This dynamically generated association threshold is used to verify the spatial differences of candidate observation data. Only observation data falling within the dynamic threshold range are retained as candidate matching data. This design effectively solves the problem of track breakage or false association caused by fixed thresholds when signal obstruction leads to sparse or drifting observation points in complex environments such as urban high-rise buildings. It achieves real-time adaptation of the screening strategy to environmental conditions.

[0042] S103: Calculate the multidimensional association cost between candidate matching data and the predicted state of the corresponding track. The multidimensional association cost integrates spatial differences, motion differences, temporal differences, and joint consistency of identity and motion.

[0043] In this embodiment, the candidate matching data after coarse screening in step S102 needs to undergo a refined similarity evaluation in this step. The multidimensional association cost is a comprehensive quantitative index used to measure the probability that a candidate observation belongs to a specific track. Among them, spatial difference, motion difference and temporal difference characterize the motion continuity of the target from three physical dimensions: position proximity, speed and heading consistency and time continuity, respectively.

[0044] In this embodiment, a joint consistency assessment of identity and motion is introduced, which breaks through the limitation of traditional methods that rely solely on identity identifiers (IDs) for matching. In practical applications, UAVs may experience temporary ID changes due to communication interference or privacy protection mechanisms. If the association is denied based solely on ID mismatch, it is easy for the same target to be incorrectly split into multiple tracks. This embodiment, by coupling identity information with motion state for judgment, can still provide a reasonable value even when IDs are inconsistent, as long as the motion characteristics are highly consistent and spatiotemporally continuous. This ensures the identity discrimination capability while taking into account the motion continuity constraint, significantly improving the robustness of association in complex electromagnetic environments.

[0045] S104: Map the multidimensional association cost to a single-frame confidence, and perform time-series recursion on the single-frame confidence based on the historical cumulative confidence to obtain the current cumulative confidence.

[0046] Step S104 in this embodiment is used to solve the problem that single-frame observation is easily affected by transient noise. Specifically, although the multidimensional association cost can reflect the matching degree of the current frame, it is an abstract cost and does not have an intuitive probabilistic meaning. Therefore, it is first mapped to a single-frame confidence level between zero and one, so that the higher the value, the more reliable the match. However, there is still a risk in making decisions based solely on the single-frame confidence level. For example, an occasional strong reflection signal may cause a false high-confidence match in a certain frame.

[0047] To address this, this embodiment further introduces a temporal recursion mechanism to fuse the current single-frame confidence score with the historical cumulative confidence score. This fusion is not a simple superposition, but rather a balancing of historical and current information through specific recursive logic: preserving the stable trend established over a long period while responding promptly to the latest observational changes.

[0048] The current cumulative confidence level obtained in this way is essentially a time-series verification result of the target attribution relationship. It can effectively filter out misjudgments caused by random impulse noise and ensure that the association decision is based on a continuous and stable chain of evidence.

[0049] S105: In response to the current cumulative confidence meeting the preset confirmation condition, confirm that the candidate matching data belongs to the corresponding track, and perform weighted fusion of the predicted state of the corresponding track and the observation state of the candidate matching data based on the single frame confidence, and update the state of the corresponding track.

[0050] The preset confirmation condition is a threshold standard for determining whether an association is valid. Only when the cumulative confidence level reaches this standard is the candidate matching data officially recognized as belonging to the corresponding track, thus avoiding the contamination of track status by low-quality associations. After confirming the association, the status update is not a simple replacement of the predicted value with the observed value, nor does it use a fixed Kalman gain. Instead, it performs adaptive weighted fusion based on the current single-frame confidence level.

[0051] Specifically, the confidence level of a single frame directly determines the weight of the observation state in the fusion process: when the observation is highly reliable, the updated track state relies more on the observation value to quickly correct trajectory deviations; when the observation is supported by accumulated evidence but the quality of a single frame is average, the updated state retains more of the inertia of the predicted value to prevent drastic trajectory changes. This adaptive update strategy of "confidence as weight" enables the track to maintain smooth continuity while sensitively tracking the target's actual maneuvers, achieving the best balance between stability and response speed.

[0052] In this embodiment, both the fixed receiving station and the mobile receiving terminal are implemented using a UAV remote identification and monitoring device. This monitoring device is specifically designed for UAV remote identification and monitoring, and can scan and receive remote identification information actively broadcast by UAVs of various brands and types. It meets the technical requirements of national standards such as GB 42590-2023 and GB 46750-2025. The device has the following functional characteristics: In terms of signal acquisition, it supports scanning of remote identification signals from multiple UAV models and multiple network transmission methods, enabling real-time communication and uploading of detection data, providing a hardware foundation for acquiring multi-source observation data in step S101; In terms of positioning capability, the device can accurately locate the position information of the UAV and the pilot, and output the position coordinates, distance, and other parameters of the UAV and the pilot in real time, providing high-precision raw observation data for subsequent trajectory association; In terms of deployment flexibility, the device is small in size and light in weight, supports multiple installation methods such as bracket type and pole type, can be directly connected to the cloud via network, or can be connected to UAV airports or ground stations for use, suitable for... The device can be deployed in various scenarios, such as rooftops and inspection vehicles in Yingcheng City. In terms of networking capabilities, multiple monitoring devices can be connected to the backend monitoring platform, expanding the coverage area through multi-device combinations to form the multi-source heterogeneous sensor network described in step S101. Regarding data management, the backend monitoring platform has data traceability capabilities, allowing users to view historical drone trajectories and other information, providing data backtracking and verification support for the accumulation of temporal evidence in step S104 and the update of flight track status in step S105. Furthermore, the device supports electronic fence settings, allowing users to customize the monitoring range size. This function, combined with the adaptive adjustment mechanism of the dynamic association threshold in this application, enables more accurate flight track association and tracking within the user-defined monitoring area.

[0053] In this embodiment, the steps of generating dynamic association thresholds and calculating multidimensional association costs are further refined. This combination of steps constitutes the core logic chain for accurately screening candidate matching data in this embodiment, aiming to solve the problems of poor adaptability of fixed thresholds and susceptibility to interference in single identity matching under complex environments.

[0054] In this embodiment, the process of generating a dynamic correlation threshold based on the signal quality parameters and time parameters corresponding to the candidate observation data includes the following sub-steps: First, the time difference between the candidate observation data and the predicted state of the corresponding track is obtained, and a time difference compensation term is generated based on the time difference, wherein the time difference compensation term is positively correlated with the time difference.

[0055] The physical implication of this design is that the longer the time interval between the observed data and the prediction time, the higher the probability that the target will maneuver or deviate from the predicted trajectory during this period. Therefore, it is necessary to automatically expand the search range to tolerate the positional uncertainty caused by time asynchrony and prevent the loss of the true target due to an excessively narrow threshold. Secondly, a signal quality compensation term is generated based on the signal quality parameters corresponding to the candidate observed data. This signal quality compensation term is negatively correlated with the signal quality parameters.

[0056] For example, when the signal-to-noise ratio decreases, it indicates that the observation data is significantly affected by multipath effects or obstruction interference, and the positional information it contains may have a large drift. In this case, relaxing the threshold can prevent noisy real observations from being misclassified as false targets. Furthermore, the historical state uncertainty of the corresponding track is obtained, and an uncertainty compensation term is generated based on this historical state uncertainty. This uncertainty compensation term is positively correlated with the historical state uncertainty.

[0057] If a track frequently experiences residual fluctuations or sparse observations in early updates, the reliability of its state estimation will decrease. In this case, more lenient matching conditions are needed to maintain tracking continuity. Finally, the preset basic threshold, time difference compensation term, signal quality compensation term, and uncertainty compensation term are superimposed to obtain the final dynamic correlation threshold.

[0058] Through this multi-dimensional compensation and superposition mechanism, the correlation threshold is no longer a static value, but a dynamic boundary that can adapt to environmental conditions and target status in real time, significantly improving the correlation robustness in harsh scenarios such as urban high-rise buildings blocking the view.

[0059] In specific implementation, the preset basic threshold can be set to 20 meters. This value is obtained by taking the 95th percentile of the planar distance distribution of adjacent frames of confirmed tracks in an open environment. The time difference compensation term is calculated by multiplying the time difference by a time difference weighting coefficient, which is calibrated to 15 meters per second. Its physical meaning is the maximum displacement that the UAV may generate per second at a typical cruise speed. The signal quality compensation term is calculated by dividing the signal-to-noise ratio weighting coefficient by the sum of the signal-to-noise ratio and the zero-prevention minimum value. The signal-to-noise ratio weighting coefficient is calibrated to 50 m·dB, and the zero-prevention minimum value is set to 0.001 to prevent division by zero errors. The calibration method of this coefficient is to statistically analyze the standard deviation of position drift under different signal-to-noise ratio conditions and establish a mapping relationship between signal-to-noise ratio and drift amount. The uncertainty compensation term is calculated by multiplying the uncertainty weighting coefficient by the standard deviation of the observation residuals of the most recent several frames of the track. The uncertainty weighting coefficient is calibrated to 2.0. This value is determined by grid search in a test scenario containing frequent maneuvers and obstructions, with the goal of minimizing the track break rate. Under the above typical parameter configurations, when the signal-to-noise ratio is 15 dB, the time difference is 0.05 seconds, and the track uncertainty is 3 meters, the dynamic threshold is approximately 20 + 15 × 0.05 + 50 / (15 + 0.001) + 2.0 × 3 ≈ 29.4 meters. However, when the signal-to-noise ratio drops sharply to 5 dB and the time difference increases to 0.2 seconds, the dynamic threshold automatically widens to approximately 20 + 15 × 0.2 + 50 / (5 + 0.001) + 2.0 × 3 ≈ 39.3 meters, demonstrating the adaptive widening capability of the threshold as observation conditions deteriorate. The generation process of the above dynamic correlation threshold can be expressed by the following formula:

[0060] in, For the track With candidate observations The dynamic correlation threshold between them, with the dimension of meters; This is a preset basic threshold, with the unit of meter, representing the maximum allowable spatial correlation range under ideal observation conditions; The time difference between the candidate observation time and the trajectory prediction time is expressed in seconds. The time difference weighting coefficient, measured in meters per second, reflects the amplification effect of time asynchrony on the threshold. The signal-to-noise ratio of the candidate observations is expressed in decibels. The signal-to-noise ratio weighting coefficient, measured in meters and decibels, reflects the amplification effect of signal quality degradation on the threshold. To prevent zero minimum values, the unit is decibel, to prevent division by zero errors when the signal-to-noise ratio is zero; For the track The historical uncertainty, in meters, is characterized by the standard deviation of the residuals from the most recent few frames of observations. This is the uncertainty weighting coefficient, which is dimensionless and reflects the amplification effect of track instability on the threshold.

[0061] Specifically, the signal quality compensation term uses a reciprocal form rather than a nonlinear subtraction form. The trend shows a nonlinear inverse relationship between receiver position drift and signal-to-noise ratio (SNR). When the SNR is high (e.g., above 15 dB), the signal quality is good, the position drift is small and insensitive to SNR changes, and the reciprocal term approaches zero, contributing very little to the threshold. When the SNR decreases (e.g., below 5 dB), the position drift increases sharply, and the reciprocal term rapidly amplifies, automatically widening the threshold to accommodate greater position uncertainty. This reciprocal form naturally depicts the monotonically increasing trend of lower SNR, larger drift, and wider threshold, and unlike linear subtraction, it does not produce negative or zero values ​​when the SNR approaches zero.

[0062] After generating the dynamic correlation threshold, this threshold is used to verify the spatial differences of candidate observation data. Specifically, the planar distance difference and altitude difference between candidate observations and the predicted track status are fused into a comprehensive spatial difference, which is then compared with the dynamic threshold. The calculation and selection criteria for the comprehensive spatial difference are as follows:

[0063] in, The difference in planar distance between the candidate observation and the predicted state is expressed in meters. The difference in height is expressed in meters. The height conversion factor is dimensionless and is used to adjust the relative weight of height difference in the overall spatial difference. When the overall spatial difference is less than or equal to the dynamic threshold, the candidate observation passes the screening and enters the subsequent scoring stage; otherwise, it is directly eliminated.

[0064] Furthermore, the reason for using weighted Euclidean distance instead of simple Euclidean distance is that in urban low-altitude surveillance scenarios, the altitude information of drones is usually measured independently by barometers or GPS, and its error characteristics differ from those of planar position. By introducing an altitude conversion factor, the weight of the altitude dimension in spatial matching can be flexibly adjusted. For example, when the altitude measurement accuracy is high, the altitude conversion factor can be increased to strengthen altitude constraints and reduce false associations between drones at different altitude levels; when the altitude measurement noise is high, the altitude conversion factor can be decreased to reduce the interference of the altitude dimension on the matching results. The typical value range of the altitude conversion factor is 0.5 to 2.0, which can be determined by statistically comparing the standard deviation of planar position error with the standard deviation of altitude error.

[0065] After filtering candidate matching data using the aforementioned dynamic association threshold, it is necessary to further calculate the multidimensional association cost between the candidate matching data and the predicted state of the corresponding track. In this process, the evaluation logic of joint consistency between identity and motion is the key feature that distinguishes this embodiment from traditional methods, specifically including the following hierarchical judgment process: When candidate matching data perfectly matches the identity identifier of the corresponding track, a first identity value is assigned. This represents the ideal matching state with the lowest cost. However, in actual low-altitude surveillance scenarios, drones may experience temporary identity changes due to unstable communication links, switching of privacy protection policies, or device restarts. To address this, when candidate matching data does not match the identity identifier of the corresponding track, but the spatial and temporal differences between them are within a preset tolerance range, the association is not directly denied. Instead, a second identity value is assigned, and this second identity value is greater than the first. This means that although the identity information is questionable, the possibility of it being the same target is still retained due to the high spatiotemporal continuity of the motion state. Only a moderate penalty is applied to reflect the risk of identity anomalies. Conversely, when candidate matching data does not match the identity identifier of the corresponding track, and the spatial or temporal differences between them exceed the preset tolerance range, it indicates that not only are the identities different, but the motion characteristics are also discontinuous. In this case, a third identity value is assigned, and this third identity value is greater than the second identity value. This higher value effectively suppresses the probability of mistakenly associating other nearby drones with the current track.

[0066] By using the value of the first identity token, the value of the second identity token, or the value of the third identity token as the cost of identity consistency in subsequent calculations, the organic integration of identity discrimination and motion continuity constraints is achieved, effectively solving the problem of track erroneous splitting caused by abnormal identity identifiers.

[0067] In practice, the first identity value is set to 0, meaning no penalty is imposed when the identities are completely identical; the second identity value is set to 0.3, and the principle for setting this value is to reflect the risk of identity anomalies (therefore it cannot be 0), but it cannot be so large that it excludes the same target with highly continuous movement (therefore it is much less than 1). Specifically, it is determined by statistically analyzing the proportion of movement continuity that is established in identity jump events. When the probability of movement continuity being established after identity jump is high, this value can be appropriately reduced to reduce the penalty for legitimate identity changes; the third identity value is set to 1, representing the maximum penalty, to effectively suppress the misassociation of other nearby drones with the current track.

[0068] The preset tolerance range in identity consistency determination includes spatial tolerance and temporal tolerance: the spatial tolerance is taken as the value of the current dynamic association threshold, that is, when the difference in planar distance between the candidate observation and the track is less than or equal to the current dynamic threshold, it is considered spatially continuous; the temporal tolerance is 1 second, that is, when the time difference between the candidate observation and the track is less than or equal to 1 second, it is considered temporally continuous. This value is determined according to the broadcast frequency of the UAV remote identification message (usually 1 Hz), ensuring that observations within two adjacent broadcast cycles can still be considered temporally continuous. The above hierarchical evaluation logic of joint identity and motion consistency can be expressed by the following formula: Furthermore, calculating the multidimensional correlation cost between candidate matching data and the predicted state of the corresponding track also includes fine-tuning the differences in each physical dimension. This involves calculating the difference in planar distance in the planar position, the difference in altitude in the vertical direction, the difference in velocity in the motion speed, and the difference in observation time between the candidate matching data and the predicted state of the corresponding track. The formulas for calculating the differences in each dimension are as follows:

[0069]

[0070]

[0071]

[0072] in, , The plane coordinates of the candidate observation; , The plane coordinates of the predicted trajectory state. Candidate observation height, Predict altitude for flight path The velocity difference between the candidate observation and the predicted trajectory state. For candidate observation speeds, Predicting speed for the flight path, The time difference between the candidate observation time and the trajectory prediction time. The corrected time for candidate observations. Predict the timing for the flight path.

[0073] It is particularly important to explain the processing logic for the heading difference: Since the heading angle is a periodic physical quantity, direct subtraction may lead to incorrect difference assessment. For example, 359 degrees and 1 degree are actually only 2 degrees different, but the arithmetic difference is 358 degrees. Therefore, this embodiment extracts the first heading angle of the candidate matching data and the second heading angle of the predicted state of the corresponding track, calculates the absolute difference between the two, and compares the absolute difference with the result of subtracting the absolute difference from a preset periodic threshold. The smaller of the two values ​​is taken as the final heading difference. The formula for calculating the heading difference is as follows:

[0074] in, The difference in heading between candidate observations and predicted track states. The heading angle for candidate observations. The heading angle for the predicted track state. The preset period threshold is set to 360 degrees. This formula compares the absolute difference with the period threshold minus the absolute difference and takes the smaller value of the two to eliminate the period jump problem of the heading angle at the boundary between 0 degrees and 360 degrees.

[0075] This minimum orbital difference processing method eliminates the interference of angle periodic jumps on cost calculation, ensuring the accuracy of heading consistency assessment. Subsequently, the plane distance difference, altitude difference, speed difference, heading difference, and time difference are normalized using corresponding normalization constants to obtain normalized difference values ​​for each dimension. The purpose of normalization is to eliminate differences in the dimensions and orders of magnitude of different physical quantities, enabling fair comparison of heterogeneous features such as position, velocity, and time on the same scale. Finally, the normalized difference values ​​for each dimension are weighted and summed with the aforementioned identity consistency cost to obtain the multidimensional association cost. The formula for calculating this total association cost is as follows:

[0076] in, For the track With candidate observations The total cost of association between them is dimensionless. , , , , , These are the weighting coefficients for planar position, altitude, speed, heading, time, and identity consistency, respectively. They are dimensionless and satisfy the following conditions: ; Simultaneously, it serves as a normalization constant for planar distance. It is a highly normalized constant with dimensions in meters. Let be the velocity normalization constant, with dimensions in meters per second. Let be the heading normalization constant, with dimensions in degrees. Let be the time normalization constant, with the dimension of seconds. As a cost of identity consistency, it is dimensionless. The differences in each dimension are divided by the corresponding normalization constant to eliminate the dimensional differences. Then, the weighted sum is obtained to obtain a unified dimensionless cost.

[0077] The necessity of normalization lies in the fact that without normalization, the numerical ranges of different physical quantities vary greatly (for example, the difference in planar distance may be tens of meters, while the time difference is only a fraction of a second). Dimensions with larger values ​​will dominate the cost value in the summation, causing smaller dimensions to be submerged and lose their ability to distinguish. The reason for using weighted summation instead of product fusion is that weighted summation allows each dimension to contribute to the total cost independently and linearly. Even if the difference in one dimension is large, the good matching of other dimensions can still effectively reduce the total cost. On the other hand, product fusion will result in the total cost being zero when the difference in any dimension is zero. It is too sensitive to single-dimensional anomalies and is not conducive to maintaining robustness in noisy environments. The reason for using weighted summation instead of the maximum value operator is that the maximum value operator only considers the worst dimension and ignores the information of other dimensions, and cannot use the consistency of multiple dimensions to enhance the confidence of judgment.

[0078] In this implementation, the typical values ​​for each normalization constant are as follows: the normalization constant for planar distance is the current dynamic correlation threshold (because it changes dynamically with the environment, the threshold value calculated in real time is used directly); the normalization constant for altitude is 120 meters (corresponding to the upper limit of urban low-altitude monitoring altitude); the normalization constant for speed is 20 meters per second (corresponding to the maximum horizontal speed of consumer-grade drones); the normalization constant for heading is 180 degrees (the maximum value of heading difference after minimum orbital difference processing); and the normalization constant for time is 1 second (corresponding to the message broadcast cycle). The typical configuration of the weight coefficients for each dimension is as follows: planar position weight 0.30, altitude weight 0.15, speed weight 0.15, heading weight 0.15, time weight 0.10, identity consistency weight 0.15, and the sum of all weights equals 1.

[0079] The calibration method for this weight combination is as follows: on the labeled test dataset, with the optimization objective of minimizing the false association rate, the optimal combination is searched in the weight space using a grid search algorithm, and then fine-tuned according to the actual scenario requirements. For example, in densely populated urban areas with tall buildings, due to significant positional drift but relatively stable headings, the planar position weight can be appropriately reduced to 0.25 and the heading weight increased to 0.20; in open airspace, where positional information is more reliable, the planar position weight can be appropriately increased to 0.35 and the heading weight reduced to 0.10. This multidimensional association cost comprehensively reflects the degree of matching between candidate observations and tracks in four dimensions: space, motion, time, and identity. The lower the cost, the higher the probability of attribution. Compared to traditional methods that rely solely on nearest neighbors, this multidimensional fusion evaluation mechanism can significantly distinguish between spatially close but differently moving targets, greatly reducing the false association rate in dense scenarios.

[0080] In this embodiment, the steps of mapping multidimensional association cost to single-frame confidence and performing temporal recursion are elaborated in detail. Specifically, the multidimensional association cost is mapped to single-frame confidence between zero and one through an exponential decay function, and the multidimensional association cost and single-frame confidence are negatively correlated. The original intention of this mapping mechanism is to solve the problem that the numerical range of multidimensional association cost is uncertain and lacks intuitive probabilistic meaning.

[0081] Since the multidimensional association cost is obtained by weighted summation of multiple normalized difference values, its theoretical range may fluctuate with weight configuration and environmental changes. It is difficult to set a unified standard for direct application in decision-making. Using an exponential decay function can compress any non-negative cost into a closed interval between zero and one, making it interpretable in a probabilistic sense. More importantly, the nonlinear characteristics of the exponential function make it more sensitive to high-quality matching: when the association cost is extremely small, the confidence level quickly approaches one, indicating a high degree of certainty; while when the association cost increases, the confidence level decreases rapidly in a nonlinear manner, thus effectively suppressing the interference of low-quality matching on subsequent decisions.

[0082] Compared to linear normalization, this transformation method better highlights the essential difference between high-quality and low-quality observations, providing a standardized input basis for subsequent time series accumulation. The mapping process of the single-frame confidence mentioned above can be expressed by the following formula:

[0083] in, For the first Candidate observations in frames With track The single-frame confidence score between values ​​is dimensionless and ranges from (0,1]. For the first The multidimensional association cost of a frame is dimensionless and ranges from [0,1]. When the confidence level approaches zero, the confidence level of a single frame approaches 1, indicating a high degree of certainty. As the association cost increases, the confidence level of a single frame decays rapidly and non-linearly, approaching zero, effectively suppressing low-quality matching.

[0084] This embodiment chooses the exponential decay function instead of the nonlinear mapping or the sigmoid function based on a comprehensive consideration of the following trend characteristics: From the perspective of value range constraints, since all normalized difference terms and weight coefficients are non-negative, the total association cost ranges from [0,1]. The value range of the exponential function is approximately [0.37,1], naturally falling within the (0,1) interval. Therefore, it can be directly used as the confidence score without additional truncation or scaling. From the perspective of decay trend, the first derivative of the exponential function near the multidimensional association cost of 0 is -1. This means that when the association cost increases from zero, the confidence score decreases at its maximum rate; while as the multidimensional association cost continues to increase, the decay rate gradually slows down. This decay characteristic of rapid initial decrease followed by slower decrease aligns with the physical intuition of association judgment: when the differences in each dimension are extremely small, the observation and track are highly matched, and the confidence score should be close to 1 and sensitive to small changes; when the differences are already large, the observation quality has already... The results are very poor. The confidence level is already low, and further changes have little impact on the decision. Compared to linear mapping, which has a constant decay rate and cannot distinguish between the slight degradation of high-quality matches and the further deterioration of low-quality matches, the cumulative confidence level will be compressed in subsequent time-series accumulation, reducing its sensitivity to high-quality evidence. Compared to the Sigmoid function, which requires additional adjustment of slope and offset parameters to adapt to different cost distributions, increasing the complexity of parameter calibration, the exponential function requires no additional parameters, is simple in form, and has deterministic behavior, making it more suitable for engineering deployment. In addition, the exponential decay function has a clear statistical meaning in probability theory: under the Gaussian noise assumption, the association cost is proportional to the logarithm of the observation likelihood probability, which is proportional to the observation likelihood probability. This gives the single-frame confidence level probabilistic interpretability under the Bayesian inference framework.

[0085] Furthermore, after obtaining the single-frame confidence, the historical cumulative confidence of the corresponding track at the previous moment is obtained, and the historical cumulative confidence and the single-frame confidence are weighted and fused using a preset forgetting factor to obtain the current cumulative confidence. The preset forgetting factor is used to balance the historical continuity and the current frame change. This step constitutes the core of the timing verification in this embodiment. In specific implementation, the calculation of the current cumulative confidence is essentially a first-order recursive filtering process.

[0086] The preset forgetting factor is usually set between 0.7 and 0.9. The choice of this parameter directly determines the trade-off between the response speed to changes in target behavior and the noise resistance. If the forgetting factor is set too high, it will rely too much on historical information. Although it can effectively smooth out instantaneous noise, the update of the cumulative confidence will be significantly delayed when the target makes a real maneuver or its identity changes legally, which may lead to missed tracking. Conversely, if the forgetting factor is set too low, it will be too sensitive to the current frame. Although the response is fast, it is easily affected by multipath reflection or sudden interference, resulting in false confirmations.

[0087] Therefore, by introducing an adjustable forgetting factor, this embodiment allows for flexible configuration of temporal memory depth based on the dynamic characteristics of the actual application scenario. For example, the forgetting factor can be appropriately increased in densely populated urban areas with high-rise buildings to enhance anti-occlusion capabilities, while it can be appropriately decreased in open airspace to improve tracking sensitivity. This temporal recursive mechanism ensures that the association decision is not based on isolated instantaneous judgments, but rather on the accumulation of evidence within a continuous time window, fundamentally improving the robustness of track association in complex environments. The recursive calculation process of the aforementioned cumulative confidence can be expressed by the following formula:

[0088] in, For the first The cumulative confidence level of a frame is dimensionless. The cumulative confidence level of the previous frame is dimensionless. The preset forgetting factor is dimensionless and has a value range of (0,1), used to balance historical continuity and current frame changes; This represents the single-frame confidence level for the current frame, and is dimensionless.

[0089] because and All are non-negative numbers and their sum is 1. Always remain within the [0,1] interval, when When the noise level is high, it places greater emphasis on historically accumulated evidence, exhibiting strong noise resistance but slow response speed; when... When the size is small, it pays more attention to changes in the current frame, and the response speed is fast but it is easily affected by momentary interference.

[0090] It should be noted that this embodiment uses a first-order recursive filtering method instead of sliding window averaging or multi-frame accumulation, based on the following design considerations: From a computational efficiency perspective, the first-order recursive form only needs to store the accumulated value of the previous frame and a forgetting factor, with constant memory overhead. It also requires only one multiplication and one addition per frame, making it suitable for real-time operation on embedded or edge computing devices. In contrast, the sliding window mean requires caching the confidence scores of the most recent N frames and updating the summation frame by frame, with memory and computational overhead increasing linearly with the window length.

[0091] From the perspective of time-weighted characteristics, the weight of each historical frame decreases exponentially, with the weight of the most recent frame being greater and the weight of the most distant frame being smaller. This exponential weighting characteristic meets the physical requirements of track association: the observations of the most recent frames best reflect the current state of the target and should be given higher weights; while the earlier frames, although still providing reference information, should have their importance decay over time.

[0092] In terms of equivalent memory depth, the equivalent window length of a first-order recursive filter is approximately... Frame, when At that time, the equivalent window length was approximately 6.7 frames, meaning that the observations of the most recent 7 frames were given primary weight, which matches the duration (3 to 7 seconds) of typical occlusion events in urban low-altitude surveillance.

[0093] Compared to simple multi-frame accumulation, the accumulation method results in a monotonically increasing confidence level over time, which cannot reflect real-time changes in the target's state and requires additional normalization processing. In contrast, the first-order recursive method naturally has a forgetting mechanism. When the target flies away or the signal continues to deteriorate, the accumulated confidence level can automatically decay, providing a basis for subsequent elimination decisions.

[0094] Furthermore, after obtaining the current cumulative confidence level, it is necessary to determine whether it meets the preset confirmation conditions. Specifically, a confirmation threshold and a rejection threshold are set, and the confirmation threshold is greater than the rejection threshold to form a state buffer. This dual-threshold design is a key improvement of this embodiment that distinguishes it from the traditional single-threshold decision method. In the traditional method, setting only a single threshold often leads to frequent state switching near the critical value, which is the so-called ping-pong effect.

[0095] For example, when signal quality fluctuates slightly around a threshold due to environmental obstruction, the track may repeatedly jump between confirmation and loss, causing track fragmentation and display jitter. To solve this problem, this embodiment constructs a hysteresis interval enclosed by a confirmation threshold and a rejection threshold. When the current cumulative confidence is greater than or equal to the confirmation threshold, the candidate matching data is confirmed to belong to the corresponding track, and a weighted fusion update of the corresponding track's state is triggered. This indicates that the chain of evidence is strong enough to safely incorporate new observations into track maintenance.

[0096] When the current cumulative confidence level is less than or equal to the elimination threshold, candidate matching data is eliminated or the association with the corresponding track is broken. This indicates that the target has likely lost tracking or an unacceptable anomaly has occurred, and losses should be stopped promptly to avoid polluting the track database. When the current cumulative confidence level is between the elimination threshold and the confirmation threshold, the candidate matching data is kept in a pending state, and the status of the corresponding track is not updated for the time being. This pending state is essentially an observation window, allowing patience when evidence is insufficient or controversial. It avoids blindly confirming or hastily discarding data, but waits for more frames of data to break the deadlock. This buffering mechanism greatly enhances the adaptability to non-stationary noise environments, effectively avoiding erroneous track deletion due to instantaneous signal quality deterioration or erroneous track creation due to chance, and ensuring the continuity and stability of track management.

[0097] In this embodiment, after confirming that the candidate matching data belongs to the corresponding track, a state update operation is performed. Specifically, based on the single-frame confidence, the predicted state of the corresponding track and the observed state of the candidate matching data are weighted and fused to update the state of the corresponding track. In this process, the single-frame confidence is determined as the observation weight, and the difference between the single-frame confidence and the observation weight is determined as the prediction weight. Subsequently, the observed state of the candidate matching data is weighted using the observation weight to obtain the first weighted state; the predicted state of the corresponding track is weighted using the prediction weight to obtain the second weighted state; finally, the first weighted state and the second weighted state are added to obtain the fused track state, so that the dependence ratio of the observed state is positively correlated with the single-frame confidence, and the dependence ratio of the predicted state is negatively correlated with the single-frame confidence. This adaptive fusion strategy of confidence as weight reflects the real-time response capability of this embodiment to dynamic changes in observation quality. Unlike the traditional Kalman filter algorithm, which usually assumes that the noise follows a Gaussian distribution and its statistical characteristics are known or change slowly, this is often difficult to hold in complex scenarios such as urban low-altitude monitoring.

[0098] In real-world environments, the error distribution of UAV remote identification signals often exhibits heavy-tailed, asymmetric, or even multimodal characteristics, and fluctuates dramatically with the degree of obstruction. Using fixed gain or adaptive gain based on model covariance can easily lead to excessive correction under strong interference, causing trajectory jumps, or insufficient correction under weak signals, resulting in trajectory drift. In contrast, this embodiment directly uses the single-frame confidence score, reflecting the matching reliability of the current frame, as the fusion weight, achieving data-driven state updates. For example, when a frame observation has passed cumulative confirmation but the single-frame confidence score is only 0.6, the observed state is automatically assigned 60% weight while retaining 40% of the predicted state inertia, thus suppressing potential measurement noise while absorbing new information. When the single-frame confidence score reaches 0.95, the observation value is almost completely trusted to achieve rapid correction. This mechanism does not require precise pre-modeling of noise statistics and can achieve smooth and accurate state estimation in non-Gaussian, non-stationary noise environments, significantly improving the accuracy and stability of trajectory updates. It should be understood that the above weighted fusion is not only applicable to linear state variables such as position and velocity, but also to periodic state variables such as heading angle. Appropriate periodization processing or smoothing logic can be applied after fusion to ensure the physical consistency of state updates. The above weighted fusion update process can be expressed by the following formula:

[0099] in, The updated track status uses the same dimensions as the individual state components (position in meters, velocity in meters per second, and heading in degrees). This is the current confirmed observation status, with the same dimensions. , To predict the state, To update the weights, which are dimensionless, the current single-frame confidence score is used. The value range is (0,1], because and All are non-negative numbers and their sum is 1, the fusion result It always lies on a convex combination between the observed and predicted states, ensuring the physical rationality of the updated state. When the confidence level of a single frame is high, As the confidence level approaches 1, the update results rely more on the observed values; when the confidence level of a single frame is low, As the value approaches zero, the updated result retains more of the inertia of the predicted value.

[0100] Furthermore, the use of a linear convex combination rather than nonlinear fusion or switching is based on the following design considerations: from the perspective of dimensional consistency, the components of the observed state and the predicted state have the same physical dimensions (position in meters, velocity in meters per second, and heading in degrees). and All are dimensionless weighting coefficients, so the dimensions of each component remain unchanged after weighted summation. The fusion result is physically meaningful. From the perspective of the boundedness of convex combinations, since... The range of values ​​ensures that the fusion result always lies between the observed and predicted values, without any extreme values ​​exceeding the range of either value. This fundamentally eliminates trajectory jumps caused by state updates.

[0101] For example, when the observed position is (100, 200) meters, the predicted position is (98, 198) meters, and the update weight is 0.7, the fusion result is (99.4, 199.4) meters, which must fall between the two. Adaptively, the update weight is directly taken from the single-frame confidence score, which is generated by a number decay function. Therefore, the fusion weight and the observation quality exhibit a monotonically increasing nonlinear mapping relationship: when the observation quality is extremely high, the update weight is close to 1, almost completely trusting the observation value to achieve rapid correction; when the observation quality is average, the update weight is in the middle range, balancing between observation and prediction; when the observation quality is poor, the update weight is close to 0, mainly relying on prediction inertia to suppress noise. This adaptive trend of trusting observations more when the evidence is stronger and preserving predictions more when the evidence is weaker allows for smooth and accurate state estimation in non-Gaussian and non-stationary noise environments without the need for precise pre-modeling of noise statistics. Compared to the matrix gain of Kalman filtering, this embodiment only requires a scalar weight to update the weight, reducing the computational complexity from matrix inversion and multiplication to scalar multiplication and addition operations, making it more suitable for real-time deployment on resource-constrained edge devices.

[0102] In this embodiment, the data preprocessing, track status prediction and smoothing, and loss of connection management mechanism are specifically elaborated. Although these steps are auxiliary support links for track association, they are the foundation for ensuring the stable operation of the core association algorithm in complex engineering scenarios such as urban high-rise buildings blocking the view, and constitute a complete data quality assurance and track lifecycle management system.

[0103] Specifically, the process of preprocessing multi-source observation data to obtain candidate observation data includes two key sub-steps: time synchronization and deduplication. Within a preset calibration window, the average deviation between the reception time and broadcast time of data from the same source is calculated. This average deviation is used to correct the reception time of subsequent multi-source observation data, outputting time-synchronized observation data. The core of this design lies in solving the problem of clock asynchrony among multi-source heterogeneous sensors. In actual deployments, there are often fixed deviations and random transmission jitter between the local clocks of different receiving stations and the UAV broadcast clock. If correction is based solely on the time difference of a single measurement, it is easily affected by instantaneous network latency or decoding errors, leading to the introduction of new noise into the correction result. Therefore, this embodiment uses a statistical mean estimation method to collect time difference samples from multiple messages within a preset calibration window (e.g., 10 seconds or 50 message cycles) and calculate the average value. This processing method can effectively filter out random jitter components and extract stable clock offsets, thereby unifying observation data from different receiving stations to the same absolute time reference. It is important to distinguish that time correction here refers to the spatiotemporal alignment of the original data, aiming to eliminate systematic errors at the data source; while time difference refers to the residual between the corrected observation time and the predicted track time, used to dynamically adjust the correlation threshold. These two processes operate at different levels and address different technical issues. The aforementioned time deviation estimation and correction process can be expressed by the following formula:

[0104] in, For receiving station The estimated time deviation, For receiving station The number of messages received within the calibration window; For receiving station The first locally recorded The time of receiving the message; For the first The broadcast timestamp of each message is used to filter out random transmission jitter and extract stable system clock offset by averaging the time differences of multiple messages within the calibration window.

[0105]

[0106] in, The corrected unified timestamp eliminates the systematic clock offset at the receiving station, unifying the observation data from different receiving stations to the same time reference.

[0107] After time synchronization, preprocessing also includes deduplication. Specifically, within a preset duplicate judgment window, multiple duplicate data from the same data source and with the same identifier are selected from the time-synchronized observation data. The signal quality parameters of each duplicate data are compared, and only the duplicate data with the highest signal quality parameter is retained as valid observation data. The remaining duplicate data are deleted, and the valid observation data is used as candidate observation data. In urban low-altitude surveillance scenarios, due to multipath effects or retransmission mechanisms of communication protocols, the same receiving station may receive multiple messages from the same UAV in a very short period of time. Without deduplication, these redundant data will cause the track update frequency to rise abnormally, and even cause state estimation oscillations due to small differences between multiple data. This embodiment chooses signal quality parameters (such as signal-to-noise ratio SNR) as the selection criterion, rather than simply retaining the earliest or latest message. The physical basis for this is that higher signal quality usually means that the message is less obstructed, reflected, or interfered with, and the position, speed, and other information it carries is closer to the true value. This data quality-based optimization strategy can improve the data purity of the input association algorithm from the source, reducing the ineffective consumption of subsequent computing resources. The above-mentioned process of duplicate detection and optimization can be expressed by the following formula:

[0108] in, and For the same receiving station The corrected times of the previous two messages, To pre-determine duplicates, and The two messages are assigned to the UAV identification number. When the time difference between the two messages is less than or equal to the duplicate determination window and the identification is the same, they are determined to be duplicate messages. Only the message with the highest signal-to-noise ratio is retained because it is least affected by obstruction and interference and has the best observation quality.

[0109] Furthermore, after updating the state of the corresponding track, the process also includes prediction and smoothing of the track state. Based on a uniform motion model, the predicted planar position and altitude for the next moment are calculated using the current velocity, heading angle, and time interval of the corresponding track. Although the UAV may maneuver during actual flight, its motion state has a high degree of continuity and inertia within the short time interval (usually about 1 second) between two adjacent observations. Using a uniform motion model for short-term extrapolation can provide a sufficiently accurate prediction benchmark while ensuring computational efficiency, providing a reasonable search center for the association matching in the next cycle. The above uniform motion prediction model can be expressed by the following formula:

[0110]

[0111]

[0112] in, and For the track The predicted planar position at the next moment. and The current planar position. The horizontal velocity at the current moment, The heading angle at the current moment. The model represents the time interval between the current observation and the next observation. It assumes that the UAV's velocity and heading remain relatively stable over a short period. The displacement increment is obtained by projecting the velocity along the x and y axes using trigonometric decomposition and multiplying it by the time interval. The predicted altitude for the next moment; The current altitude. This represents the current rate of ascent or descent.

[0113] For the special state variable of heading angle, the historical heading angles of the most recent consecutive frames of the corresponding track are extracted, and the sum of the sine and cosine values ​​of each historical heading angle are calculated. The sum of the sine and cosine values ​​is then calculated using a two-dimensional arctangent function to obtain the smoothed heading angle, in order to eliminate the periodic jumps in angle. The heading angle is a physical quantity with a periodicity of 360 degrees, and directly averaging it will lead to serious logical errors.

[0114] For example, when the heading angles of two consecutive frames are 359 degrees and 1 degree respectively, their arithmetic mean is 180 degrees, which is completely opposite to the actual flight direction (near true north). The vector synthesis smoothing algorithm used in this embodiment essentially converts the angle into vectors on the unit circle, superimposes them, and then restores the angle through arctangent operation. In specific implementation, the smoothing window takes the heading angles of the three most recent consecutive frames. The length of this window is set based on the following: when the message broadcast frequency is 1 Hz, three frames correspond to a time span of about 3 seconds. Within this time window, the heading change of the UAV usually does not exceed tens of degrees, which is sufficient to ensure the directional representativeness of the vector synthesis result; at the same time, the three-frame window is short enough not to smooth out the real heading maneuver due to an excessively large span. If the window is too short (for example, only one frame is taken), it will not have a smoothing effect; if the window is too long (for example, more than 10 frames), it may also smooth out the real heading change, resulting in an increase in prediction deviation. This method naturally bridges the boundary discontinuity between 0 degrees and 360 degrees, ensuring a smooth result that conforms to physical facts under any heading change, and avoiding trajectory display anomalies or prediction deviations caused by angle jumps. The above heading angle vector synthesis smoothing algorithm can be expressed by the following formula:

[0115] in, The smoothed heading angle, For the track In the most recent N consecutive frames (N=3 in this example), the first The heading angle before the frame, atan2 is a two-dimensional arctangent function, whose inputs are the sum of the sine and cosine values ​​of the heading angles of all historical frames. This formula converts the heading angles of each frame into vectors on the unit circle and then superimposes them, and then restores them to angles through arctangent operation, naturally crossing the boundary discontinuity between 0 degrees and 360 degrees.

[0116] To address the unavoidable signal loss in complex environments, this embodiment also incorporates a robust loss-of-connection management mechanism. Specifically, when no candidate matching data meeting the dynamic association threshold and preset confirmation conditions is found within the current time period, the corresponding track is marked as lost, and the consecutive lost-of-connection frame count is incremented. This indicates that at the current moment, no observation data can be attributed to the track with sufficient confidence. However, this does not immediately mean the target has disappeared. In densely populated urban areas with tall buildings, temporary building obstructions or signal blind spots are common causes of observation interruptions. If tracks are deleted immediately upon loss of connection, a large number of originally continuous tracks will be fragmented, severely impacting monitoring effectiveness. Therefore, it is determined whether the consecutive lost-of-connection frame count exceeds a preset loss-of-connection limit. If the consecutive lost-of-connection frame count does not exceed the preset loss-of-connection limit, the predicted state of the corresponding track is maintained, and it continues to participate in subsequent associations.

[0117] During this period, although no new observations are updated on the track, it still relies on motion models for pure prediction and extrapolation to maintain its survivability, waiting for the target to reappear in the field of vision so that it can be quickly resumed. Conversely, if the number of consecutive lost frames exceeds the preset loss limit, the corresponding track is terminated and associated resources are released. The setting of this limit (e.g., 5 or 10 frames) is a compromise between preventing track breakage and preventing false alarms. If the retention time is too short, it will not be enough to cover the typical occlusion duration. If the retention time is too long, it may cause targets that have flown away or landed to occupy resources in the form of ghost tracks or even interfere with the establishment of new targets. Through this limited-time loss retention mechanism, this embodiment ensures track continuity while also maintaining the effectiveness of resources and the cleanliness of the track database. The above loss management mechanism can be expressed by the following formula:

[0118] in, For the track The number of consecutive lost frames is incremented by one when no candidate matching data that meets the dynamic association threshold and preset confirmation conditions is found within the current time period. When the number of lost frames does not exceed the preset loss limit, the trajectory prediction state is maintained and it continues to participate in subsequent associations. When the number of lost frames exceeds the preset loss limit, the trajectory is terminated and the association resources are released.

[0119] In this embodiment, in order to more clearly illustrate the feasibility and beneficial effects of the technical solution in actual engineering scenarios, a comprehensive application verification is carried out using an urban low-altitude monitoring scenario as an example. In this scenario, the target area is the densely populated urban center, and the multi-source observation data is specifically represented by Remote ID messages broadcast by UAVs. These messages are jointly collected by fixed receiving stations deployed on the rooftops of high-rise buildings and mobile receiving terminals mounted on ground inspection vehicles.

[0120] Each Remote ID message contains fields such as UAV identification number, broadcast timestamp, receive timestamp, planar position, altitude, horizontal speed, heading angle, signal strength, and signal-to-noise ratio (SNR). The UAV identification number serves as an identifier, signal strength and SNR are signal quality parameters, and the broadcast and receive timestamps constitute the time parameters. Due to the urban canyon effect, observation data in this scenario generally suffers from severe multipath reflections, frequent signal blockage, and clock asynchrony between different receiving stations, making it a typical test environment for verifying the robustness of the trajectory association method in this embodiment.

[0121] In this embodiment, the key parameters are configured as follows: the preset calibration window is set to 10 seconds to statistically analyze the time deviation of the receiving station; the duplicate detection window is set to 500 milliseconds to identify duplicate messages from the same station; the confirmation threshold is set to 0.8, and the rejection threshold is set to 0.3, forming a state buffer with a width of 0.5; the preset disconnection limit is set to 5 frames, corresponding to approximately 5 seconds of pure prediction retention time; the forgetting factor is set to 0.85 to balance historical inertia and new observation response. In the multidimensional association cost calculation, the weights of each dimension are configured as follows: planar position weight 0.30, altitude weight 0.15, velocity weight 0.15, heading weight 0.15, time weight 0.10, and identity consistency weight 0.15; the normalization constants are configured as follows: altitude normalization constant 120 meters, velocity normalization constant 20 meters per second, heading normalization constant 180 degrees, and time normalization constant 1 second; the three grading values ​​of the identity consistency cost are 0, 0.3, and 1, respectively. In the dynamic association threshold generation, the preset base threshold is set to 20 meters, the time difference weighting coefficient is set to 15 meters per second, the signal-to-noise ratio weighting coefficient is set to 50 meters per decibel, the uncertainty weighting coefficient is set to 2.0, and the minimum zero protection value is set to 0.001. These parameters are not arbitrarily chosen, but are derived from a comprehensive trade-off between the typical flight speed of urban low-altitude UAVs (10-20 meters per second) and the broadcast frequency of Remote ID messages (1 Hz). The calibration methods for each parameter have been detailed in the previous section on parameter calibration standardization, and will not be repeated here. For example, a 10-second calibration window is sufficient to cover dozens of message samples to eliminate random jitter, while a 5-frame disconnection limit can tolerate short-term signal interruptions caused by most building obstructions, while preventing targets that have flown away from occupying resources for extended periods.

[0122] The following simulation demonstrates the entire process of a drone flying from an open area into a high-rise building's obstruction zone and then out again, showcasing the complete data flow logic of the method in this embodiment. First, in the preprocessing stage of step S101, raw Remote ID messages from multiple receiving stations are received. For a specific receiving station, within a 10-second calibration window, its received timestamp is calculated to be on average 80 milliseconds later than its broadcast timestamp. This deviation is then used to correct the reception time of all subsequent messages from that station, achieving a unified time base for multi-source data. Next, within a 500-millisecond duplication detection window, two messages with the same drone identification number but signal-to-noise ratios (SNRs) of 15dB and 8dB respectively are detected at the same receiving station. Based on the selection principle, only the message with an SNR of 15dB is retained as candidate observation data, effectively avoiding redundant data interference caused by multipath reflection.

[0123] Please see Figure 2 , Figure 2 This application provides a schematic diagram of a curve associated with a tracking track. Figure 2The tracking results of this embodiment are demonstrated. Throughout the entire process of the UAV flying from an open area into a blocked area and then out again, the flight path remains single and continuous. In the blocked area, this embodiment uses a dynamic threshold relaxation and disconnection maintenance mechanism, along with motion model prediction and extrapolation, to ensure that although the flight path deviates to some extent, it does not break. After flying out of the blocked area, accurate tracking is quickly restored.

[0124] The process then proceeds to the dynamic filtering stage in step S102. When the drone flies into the shadow area of ​​a tall building, the signal-to-noise ratio (SNR) of the message received by a certain receiving station drops sharply from 15 dB to 5 dB. At this point, if a traditional fixed threshold (e.g., 20 meters) is used, the observation point may be discarded due to its location drift exceeding the range. However, in this embodiment, a larger signal quality compensation term is generated based on the reduced SNR. Simultaneously, considering the 200-millisecond asynchronous delay between the observation time and the prediction time at this moment, a corresponding time difference compensation term is superimposed, automatically widening the final dynamic association threshold from 20 meters to 35 meters. This adaptive adjustment successfully incorporates noisy real observations into the candidate matching dataset, preventing premature track loss due to environmental degradation.

[0125] In the multidimensional association cost calculation in step S103, a typical identity anomaly event occurred: the UAV's broadcast UAV identification number temporarily changed from "UAV_001" to "UAV_001_B" due to a communication link switch. In traditional methods, this would be directly identified as a new target. However, the evaluation logic in this embodiment detected that although the identity identifiers were inconsistent, the difference in planar distance between the candidate matching data and the predicted trajectory state was only 3 meters, and the time difference was only 50 milliseconds, both within the preset tolerance range. Therefore, a second identity value (medium penalty) was assigned instead of a third identity value (high penalty). Combining the normalized planar distance difference, altitude difference, speed difference, and heading difference processed by minimum orbital difference, the calculated multidimensional association cost was still low, indicating that although the identity was questionable, the strong continuity of the motion state supported the assumption that they were the same target.

[0126] Next, in step S104, the multidimensional association cost is mapped to a single-frame confidence score of 0.75 using an exponential decay function. Since 0.75 is below the confirmation threshold of 0.8 but above the rejection threshold of 0.3, the association is not immediately confirmed; instead, it is placed in a pending state. A forgetting factor of 0.85 is used to recursively fuse the historical cumulative confidence score of 0.9 from the previous time step, resulting in a current cumulative confidence score of 0.8775. Because 0.8775 is greater than the confirmation threshold of 0.8, the candidate matching data is formally confirmed to belong to the corresponding track at the next time step. This process clearly demonstrates how the dual-threshold buffer and the multi-frame accumulation mechanism work together: neither is it hastily rejected due to a single-frame identity anomaly, nor is it blindly trusted because a single match is acceptable; instead, robust decision-making is achieved through a temporal chain of evidence.

[0127] Finally, in the state update step S105, given that the current single-frame confidence is 0.75, the observation weight is set to 0.75 and the prediction weight is set to 0.25, and a weighted fusion is performed on the track state. This means that 75% of the updated position comes from the latest observation and 25% comes from the model prediction, which corrects the trajectory deviation and suppresses abrupt jumps caused by observation noise. At the same time, for the heading angle, the sine and cosine vector sum of the most recent 3 frames is used for smoothing, eliminating the display jumps that may occur when the angle crosses the 0-degree boundary. When the UAV completely enters the blind zone, resulting in 3 consecutive frames without matching, the disconnection management mechanism is triggered to maintain the track prediction state and accumulate the disconnection count. When the UAV flies out of the blind zone and is recaptured, the track can be seamlessly continued without any break because the disconnection count has not exceeded the 5-frame limit.

[0128] Please see Figure 3 and Figure 4 , Figure 3 This is a schematic diagram showing the costs of various dimensions in a typical scenario of temporary identity change provided in the embodiments of this application. Figure 4 This is a schematic diagram of the confidence change curve in a typical scenario of temporary identity change provided in this application embodiment. During the identity change interval of frames 11-20, the identity cost suddenly increases from 0 to 0.3, resulting in a corresponding increase in the total association cost. However, since the costs of motion dimensions such as position, speed, altitude, and heading remain low (motion state is continuous), the increase in the total cost is controlled within a reasonable range. Correspondingly, the single-frame confidence decreases from about 0.95 to about 0.70 during the identity change. However, since 0.70 is still much higher than the elimination threshold of 0.3, and the cumulative confidence remains above 0.85 due to historical inertia (higher than the confirmation threshold of 0.8), the track association can be maintained. After frame 21, the identity returns to normal, and all costs and confidence quickly return to normal levels. This process demonstrates the core value of the joint consistency assessment of identity and motion in this embodiment. Even if the identity is abnormal, as long as the motion features are highly consistent and spatiotemporally continuous, a reasonable cost can still be given, avoiding the problem of existing methods incorrectly splitting the same target into multiple tracks due to ID mismatch.

[0129] Please see Figure 5Figure 5 is a schematic diagram of the typical confidence curve of the UAV passing through the open area, the obstructed area and the recovery area provided in the embodiment of this application. It simulates the complete process of a UAV flying from an open area into a high-rise building obstructed area and then flying out again, showing the trajectory of the change of single-frame confidence and cumulative confidence. Three obvious stages can be observed in the figure: (1) Open area (frames 1-15): the single-frame confidence is stable at about 0.85, the cumulative confidence rises rapidly and stabilizes above 0.85, always higher than the confirmation threshold, and the flight path is maintained normally; (2) Obstructed area (frames 15-35): the single-frame confidence drops to about 0.5 due to signal deterioration, but the cumulative confidence rises due to the loss of signal. Due to the historical inertia effect of forgetting factor λ=0.85, the rate of decline is significantly slower than that of single-frame confidence. During frames 18-22, it remains above the confirmation threshold, successfully maintaining the track association. This demonstrates the anti-occlusion capability of the time-series recursive mechanism. Even if the single-frame observation quality temporarily declines, the historical accumulated trust can still support the association decision. (3) Recovery area (frames 35-60): The single-frame confidence recovers, and the accumulated confidence quickly recovers to a high level. This process proves that the time-series verification mechanism implemented by first-order recursive filtering in this embodiment can effectively filter out misjudgments caused by random impulse noise and ensure that the association decision is based on a continuous and stable chain of evidence.

[0130] In contrast, parallel tests were conducted using the traditional method of "fixed distance threshold (20 meters) + single-frame identity verification" under the same test data. The results showed that when the signal-to-noise ratio decreased as the UAV entered an obstructed area, the traditional method lost 40% of effective observation points due to its narrow threshold; at the instant the identity identifier changed, the traditional method directly terminated the original track and created an incorrect track; during brief signal interruptions, the traditional method lacked a mechanism to maintain connection, resulting in the track breaking into three segments. In contrast, the track using the proposed solution maintained a single, continuous ID identifier throughout the flight, improving track smoothness by over 60%, and achieving 98% tracking integrity in obstructed areas. This comparative result strongly demonstrates that the proposed solution, through dynamic threshold adaptation, identity-motion joint fault tolerance, temporal confidence accumulation, and adaptive fusion update mechanisms, effectively solves the problems of track misassociation and track breakage caused by multi-source asynchronous operation, signal fluctuations, and identity anomalies in complex urban environments, demonstrating significant engineering practical value.

[0131] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for associating UAV tracks based on multi-source remote identification signals, characterized in that, include: Acquire multi-source observation data within the target area and perform preprocessing on the multi-source observation data to obtain candidate observation data; Based on the signal quality parameters and time parameters corresponding to the candidate observation data, a dynamic correlation threshold is generated, and the candidate matching data that meets the spatial difference requirements is selected from the candidate observation data using the dynamic correlation threshold. Calculate the multidimensional association cost between candidate matching data and the predicted state of the corresponding track. The multidimensional association cost integrates spatial differences, motion differences, temporal differences, and joint consistency of identity and motion. The multidimensional association cost is mapped to a single-frame confidence score, and the single-frame confidence score is recursively calculated based on the historical cumulative confidence score to obtain the current cumulative confidence score. In response to the current cumulative confidence meeting the preset confirmation conditions, the candidate matching data is confirmed to belong to the corresponding track, and the predicted state of the corresponding track and the observed state of the candidate matching data are weighted and fused based on the single frame confidence, and the state of the corresponding track is updated.

2. The method according to claim 1, characterized in that, Based on the signal quality parameters and time parameters corresponding to the candidate observation data, a dynamic correlation threshold is generated, including: The time difference between candidate observation data and the predicted state of the corresponding track is obtained, and a time difference compensation term is generated based on the time difference. The time difference compensation term is positively correlated with the time difference. A signal quality compensation term is generated based on the signal quality parameters corresponding to the candidate observation data. The signal quality compensation term is negatively correlated with the signal quality parameters. The historical state uncertainty of the corresponding track is obtained, and an uncertainty compensation term is generated based on the historical state uncertainty. The uncertainty compensation term is positively correlated with the historical state uncertainty. The preset basic threshold, time difference compensation term, signal quality compensation term, and uncertainty compensation term are superimposed to obtain the dynamic correlation threshold.

3. The method according to claim 1, characterized in that, When calculating the multidimensional association cost between candidate matching data and the predicted state of the corresponding track, the evaluation logic for joint identity and motion consistency includes: When the candidate matching data is completely consistent with the identity identifier of the corresponding track, the first identity value is assigned. When the candidate matching data is inconsistent with the identity of the corresponding track, and the spatial and temporal differences between the two are within the preset tolerance range, a second identity value is assigned, and the value of the second identity value is greater than that of the first identity value. When the candidate matching data is inconsistent with the identity of the corresponding track, and the spatial or temporal difference between the two exceeds the preset tolerance range, a third identity value is assigned, and the value of the third identity value is greater than that of the second identity value. The value of the first identity, the value of the second identity, or the value of the third identity are used as identity consistency costs to participate in the calculation of multidimensional association costs.

4. The method according to claim 3, characterized in that, Calculating the multidimensional association cost between candidate matching data and the predicted state of the corresponding track also includes: Calculate the difference in planar distance between the candidate matching data and the predicted state of the corresponding track, the difference in height in the vertical direction, the difference in velocity in the movement speed, and the difference in time in the observation time. Extract the first heading angle of the candidate matching data and the second heading angle of the predicted state of the corresponding track, calculate the absolute difference between the first heading angle and the second heading angle, and compare the absolute difference with the result of subtracting the absolute difference from the preset period threshold, and take the smaller value of the two as the heading difference; The plane distance difference, altitude difference, speed difference, heading difference, and time difference are normalized using the corresponding normalization constants to obtain the normalized difference values ​​for each dimension. The multidimensional association cost is obtained by weighted summation of the normalized difference values ​​of each dimension and the identity consistency cost.

5. The method according to claim 4, characterized in that, The multidimensional association cost is mapped to a single-frame confidence score, and a time-series recursion is performed on the single-frame confidence score based on the historical cumulative confidence score to obtain the current cumulative confidence score, including: The multidimensional association cost is mapped to a single-frame confidence level between zero and one by using an exponential decay function. The multidimensional association cost and the single-frame confidence level are negatively correlated. Obtain the historical cumulative confidence level of the corresponding track at the previous moment; The historical cumulative confidence and single-frame confidence are weighted and fused using a preset forgetting factor to obtain the current cumulative confidence. The preset forgetting factor is used to balance historical continuity and current frame changes.

6. The method according to claim 5, characterized in that, In response to the current cumulative confidence level meeting the preset confirmation conditions, the candidate matching data is confirmed to belong to the corresponding track, including: Set an acceptance threshold and an rejection threshold. If the acceptance threshold is greater than the rejection threshold, a status buffer will be formed. When the current cumulative confidence level is greater than or equal to the confirmation threshold, the candidate matching data is confirmed to belong to the corresponding track, and the weighted fusion is triggered to update the status of the corresponding track. When the current cumulative confidence level is less than or equal to the elimination threshold, candidate matching data is eliminated or the association with the corresponding track is broken; When the current cumulative confidence level is between the elimination threshold and the confirmation threshold, the pending state of the candidate matching data is maintained, and the status of the corresponding track is not updated for the time being.

7. The method according to claim 6, characterized in that, Based on the single-frame confidence score, a weighted fusion is performed between the predicted state of the corresponding track and the observed state of the candidate matching data to update the state of the corresponding track, including: The confidence level of a single frame is determined as the observation weight, and the difference between the observation weight and the prediction weight is determined as the prediction weight. The observation states of the candidate matching data are weighted using observation weights to obtain the first weighted state; The predicted states of the corresponding trajectories are weighted using the prediction weights to obtain the second weighted state; The first weighted state is added to the second weighted state to obtain the fused track state, which makes the dependence ratio of the observed state positively correlated with the single-frame confidence, and the dependence ratio of the predicted state negatively correlated with the single-frame confidence.

8. The method according to claim 1, characterized in that, Multi-source observation data undergoes preprocessing to obtain candidate observation data, including: Within the preset calibration window, the average deviation between the reception time and the broadcast time of the same data source is calculated. The average deviation is used to correct the reception time of subsequent multi-source observation data, and the time-synchronized observation data is output. Within the preset duplicate detection window, multiple duplicate data entries from the same data source and with the same identity are filtered out from the time-synchronized observation data. Compare the signal quality parameters of multiple duplicate data points, retain only the duplicate data point with the highest signal quality parameter as the valid observation data, delete the remaining duplicate data, and use the valid observation data as the candidate observation data.

9. The method according to claim 8, characterized in that, After updating the status of the corresponding track, the method also includes: Based on the uniform motion model, the predicted plane position and predicted altitude at the next moment are calculated using the current speed, heading angle and time interval of the corresponding track. Extract the historical heading angles of the most recent consecutive frames of the corresponding track, and calculate the sum of the sine and cosine values ​​of each historical heading angle respectively; The smoothed heading angle is obtained by calculating the sum of the sine and cosine values ​​using a two-dimensional arctangent function, thus eliminating periodic angle jumps.

10. The method according to claim 9, characterized in that, If no candidate matching data that meets the dynamic association threshold and preset confirmation conditions is found within the current time period, the corresponding track will be marked as lost and the number of consecutive lost frames will be incremented by one. Determine if the number of consecutive lost connections exceeds the preset loss limit; If the number of consecutive lost frames does not exceed the preset loss limit, the prediction status of the corresponding track will be maintained and it will continue to participate in subsequent correlations. If the number of consecutive lost connections exceeds the preset loss limit, the corresponding flight path will be terminated and associated resources will be released.

Images