Unmanned aerial vehicle data processing method and system based on kinematic constraints and computer

Abstract

The present application relates to a method and system for processing data of unmanned aerial vehicles based on kinematic constraints, and a computer, the method comprising receiving raw data of non-cooperative unmanned aerial vehicle perception positioning from at least two heterogeneous sensors; performing coordinate conversion and spatial decoupling on the raw data to obtain a set of data components in multiple orthogonal directions; performing linearization processing on the set of data components based on a preset sliding space-time window, and applying instantaneous velocity in each direction as a kinematic constraint to identify and filter out abnormal data components violating the kinematic constraint; performing first-stage interpolation repair on data gaps; dividing the overall data stream after the first-stage interpolation repair into multiple time windows, evaluating the continuity of the data stream according to the data density of each time window, and performing second-stage interpolation repair on the data stream of the time window evaluated as weakly continuous; the present application guarantees the spatiotemporal coherence and reliability of the reconstructed data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method, system, and computer for processing UAV data based on kinematic constraints. Background Technology

[0002] With the increasing prevalence of drone applications, the need for effective perception and localization of non-cooperative drones (i.e., unauthorized or unidentified aircraft) is growing. In practical applications, especially in complex environments such as cities, drone data acquisition is often severely interfered with by factors such as multipath effects and signal blockage. This results in a large number of anomalies and data interruptions in the measurement trajectories reported by sensors (such as radar, Time Difference of Origin Positioning System (TDOA)).

[0003] Existing technologies primarily rely on statistical methods to detect and remove outliers from raw measurements, which has the following significant drawbacks: Spatial dimensional coupling problem: The differences in the horizontal and vertical motion of the UAV were not fully considered, and the observation information from different spatial dimensions was not effectively decoupled. During processing, information with discriminative value in different spatial dimensions is easily removed along with the data and is difficult to retain.

[0004] Lack of prior physical constraints: Relying solely on statistical methods without incorporating the drone's own motion performance characteristics (such as maximum speed and acceleration limits) as physical constraints leads to the misjudgment of the drone's actual sudden movements (such as sharp turns) as noise, resulting in the loss of crucial information about the flight path.

[0005] Impaired data continuity: Existing methods primarily rely on outlier removal, leading to a loss of temporal continuity in the preprocessed data. The lack of effective data continuity restoration strategies introduces additional systematic errors into subsequent data processing (such as trajectory tracking and data fusion).

[0006] Therefore, there is an urgent need for a data processing mechanism that can integrate physical constraints, effectively decouple multidimensional information, and maintain data continuity. Summary of the Invention

[0007] To address the technical problems in existing technologies, such as dimensional coupling, lack of physical constraints, and poor data continuity, when processing multimodal perception data from non-cooperative UAVs, this invention provides a UAV data processing method, system, and computer based on kinematic constraints.

[0008] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: The method for UAV data processing based on kinematic constraints includes the following steps: Receive raw data from non-cooperative UAV perception and localization from at least two heterogeneous sensors; The original data is subjected to coordinate transformation and spatial decoupling to obtain a set of data components in multiple orthogonal directions; The data component set is linearized based on a preset sliding spatiotemporal window, and instantaneous velocities are applied in each direction as kinematic constraints to identify and filter out abnormal data components that violate the kinematic constraints. The first stage of interpolation repair is performed on the data gaps caused by filtering out the abnormal data components; The overall data stream after the first stage of interpolation repair is divided into multiple time windows. The continuity of the data stream in each time window is evaluated based on the data density. The data stream in the time window that is evaluated as having weak continuity is subjected to the second stage of interpolation repair to enhance the spatiotemporal coherence of the overall data stream.

[0009] The beneficial effects of this invention are as follows: This invention uses kinematic constraints as physical prior knowledge, orthogonally decouples the spatiotemporal dimension characteristics of measurement data, and combines window thresholds to process interrupted data continuously. While effectively separating the actual flight path of the UAV, it ensures the spatiotemporal coherence and credibility of the reconstructed data.

[0010] Furthermore, the original data undergoes coordinate transformation and spatial decoupling to obtain a set of data components in multiple orthogonal directions, specifically including the following steps: Obtain the original data points of the original data; wherein, the original data points include geographic coordinates and timestamps; Select a preset reference origin, and transform the geographic coordinates into positional components in the east, north, and sky directions under the station center coordinate system through a coordinate transformation matrix; The transformed location components are sorted chronologically based on the timestamps to form an ordered set of data components.

[0011] Furthermore, identifying and filtering out anomalous data components that violate the kinematic constraints specifically includes the following steps: Within the sliding spacetime window, for each orthogonal direction, calculate the instantaneous velocity sequence of the data component in that direction; For the instantaneous velocity sequence in each direction, the box plot algorithm is executed to determine the theoretical normal range of instantaneous velocity values ​​in that direction; If the instantaneous velocity of a data point in any direction exceeds the theoretical normal range for that direction, then the component of the data point in that direction is determined to violate the kinematic constraints, and a low-precision label is assigned to that component. Data components with low-precision labels are filtered out as null values, and the steps of this claim are iteratively executed until all data has been processed.

[0012] Furthermore, it also includes the following steps: By combining the data source information and precision labels of the data points, a binary confidence label is constructed for each data point after the interpolation repair is completed in the first and second stages; wherein, the binary confidence label is used to characterize the credibility of the data point for decision-making in subsequent data processing stages.

[0013] Furthermore, the method for constructing binary confidence labels is as follows: If all dimensional components of a data point are directly measured by a sensor and are all marked as high precision, then it is determined to be a high-confidence data point. If any dimension component of a data point originates from the interpolation generation in the second stage, or if any dimension component is marked as low precision, it is determined to be a low-confidence data point.

[0014] Furthermore, the interpolation repair method in the first stage adopts the linear time-proportional interpolation method; The linear time-proportional interpolation method is as follows: For any data gap in an orthogonal direction, obtain the two adjacent data points that have undergone spatial decoupling before and after the time point. Based on the time and location information of two spatially decoupled data points, interpolated data is generated to fill the data gaps through linear scaling.

[0015] Furthermore, a second stage of interpolation is performed on the weakly continuous window, specifically as follows: Set a data density threshold; For a given time window, calculate the density of data points within it; If the data point density is lower than the data density threshold, the window is determined to be a weakly continuous window, and the midpoint of the time domain of the window is taken as the time reference, and interpolation is performed based on the data of the surrounding windows. If the data point density is not lower than the data density threshold, then the window is determined to be a strongly continuous window, and its original data is retained.

[0016] Furthermore, the length of the sliding spatiotemporal window is predetermined to be optimal through sensitivity analysis; The sensitivity analysis includes the following steps: Set different window lengths; Calculate the root mean square error between the denoised data and the reference data with different window lengths; Based on the trend of the root mean square error, the window length corresponding to when the root mean square error tends to stabilize and is close to its minimum value is selected as the optimal length.

[0017] To address the aforementioned technical problems, this invention provides a UAV data processing system based on kinematic constraints, the specific technical content of which is as follows: A kinematically constrained UAV data processing system includes: A data acquisition module is used to receive raw data from non-cooperative UAV perception and positioning from at least two heterogeneous sensors; The data denoising module is used to perform coordinate transformation and spatial decoupling on the original data to obtain a set of data components in multiple orthogonal directions; the set of data components is linearized based on a preset sliding spatiotemporal window, and instantaneous velocity is applied as kinematic constraint in each direction to identify and filter out abnormal data components that violate the kinematic constraint. The data interpolation and repair module is used to perform a first-stage interpolation and repair on the data gaps formed by filtering out the abnormal data components; the overall data stream after the first-stage interpolation and repair is divided into multiple time windows, the continuity of the data stream in each time window is evaluated based on the data density, and a second-stage interpolation and repair is performed on the data stream in the time window that is evaluated as having weak continuity, so as to enhance the spatiotemporal coherence of the overall data stream.

[0018] To solve the above-mentioned technical problems, the present invention provides a computer, the specific technical content of which is as follows: A computer includes a memory and one or more processors, wherein executable code is stored in the memory, and when the one or more processors execute the executable code, the steps of the above-described kinematic constraint-based UAV data processing method are implemented. Attached Figure Description

[0019] Figure 1 This is a flowchart of a UAV data processing method based on kinematic constraints in an embodiment of the present invention; Figure 2 This is a framework diagram of UAV data processing based on kinematic constraints in an embodiment of the present invention; Figure 3 This is a flowchart of the staged interpolation strategy in an embodiment of the present invention; Figure 4 This is a schematic diagram of the experimental area and experimental equipment in an embodiment of the present invention; Figure 5 This is a diagram showing the device's sensing range in an embodiment of the present invention; Figure 6 This is a multi-scenario flight track information map in an embodiment of the present invention; Figure 7 This is a test height curve for a 120m scene in an embodiment of the present invention; Figure 8 This is a test height curve for a 400m scene in an embodiment of the present invention; Figure 9 This is a test height curve for a 500m scene in an embodiment of the present invention; Figure 10This is a diagram illustrating the measurement data of the heterogeneous sensor in an embodiment of the present invention; Figure 11 This is a graph showing the noise reduction experimental data from a 120m measurement in an embodiment of the present invention. Figure 12 This is a graph showing the noise reduction experimental data from a 400m measurement in an embodiment of the present invention. Figure 13 This is a graph showing the noise reduction experimental data from a 500m measurement in an embodiment of the present invention. Figure 14 This is a trajectory curve diagram after data reconstruction in an embodiment of the present invention. Detailed Implementation

[0020] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0021] like Figure 1 and Figure 2 As shown, this embodiment provides a UAV data processing method based on kinematic constraints, including the following steps: S1. Receive raw data from non-cooperative UAV perception and localization from at least two heterogeneous sensors; the raw data includes radar data and TDOA system data.

[0022] S2. Perform coordinate transformation and spatial decoupling on the original data to obtain a set of data components in multiple orthogonal directions; The original data is subjected to coordinate transformation and spatial decoupling to obtain a set of data components in multiple orthogonal directions, specifically including the following steps: Obtain the original data points of the original data; wherein, the original data points include geographic coordinates and timestamps; A preset reference origin is selected, and the geographic coordinates are transformed into positional components in the east, north, and celestial directions under the station-centered coordinate system using a coordinate transformation matrix. These positional components correspond to the longitude, latitude, and altitude components under the WGS-84 geodetic coordinate system. The east, north, and celestial directions are represented by the E, N, and U directions, respectively.

[0023] The transformed location components are sorted chronologically based on the timestamps to form an ordered set of data components.

[0024] By constructing an ENU coordinate system and mapping the original data from the WGS-84 geodetic coordinate system to this ENU coordinate system, subsequent calculations can be facilitated. Specifically, assuming in the time domain... ( , Represents the timestamp number. and heterogeneous sensors The set of reported measurement data It can be represented as (1) In the formula, , These respectively represent data sources such as radar data and heterogeneous sensor system data, i.e., TDOA system data; Representing the The position vector of each measurement data point , , These represent longitude, latitude, and altitude information, respectively.

[0025] The measurement points are converted from WGS-84 geodetic coordinates to station center coordinates using a standard conversion process. Switch to the Flight Experiment Center area Using the station-centered coordinate system with reference to the origin, its northeast-to-central coordinates are obtained. The conversion process expression is as follows: (2) In the formula, For reference coordinates The transformation matrix is ​​used to convert the differences in longitude, latitude, and altitude into east, north, and celestial displacements in the station center coordinate system; , , They are respectively Based on in all directions The transformed east, north, and celestial coordinate components. Simultaneously, the transformed data points. With the corresponding set of data components It can be represented as follows: .

[0026] S3. Linearize the data component set based on a preset sliding spatiotemporal window, and apply instantaneous velocity as kinematic constraint in each direction to identify and filter out abnormal data components that violate the kinematic constraint. Identifying and filtering out anomalous data components that violate the kinematic constraints specifically includes the following steps: Within the sliding spacetime window, for each orthogonal direction, calculate the instantaneous velocity sequence of the data component in that direction; For the instantaneous velocity sequence in each direction, the box plot algorithm is executed to determine the theoretical normal range of instantaneous velocity values ​​in that direction; If the instantaneous velocity of a data point in any direction exceeds the theoretical normal range for that direction, then the component of the data point in that direction is determined to violate the kinematic constraints, and a low-precision label is assigned to that component. Data components with low-precision labels are filtered out as null values, and the steps of this claim are iteratively executed until all data has been processed.

[0027] The length of the sliding spatiotemporal window is predetermined to be optimal through sensitivity analysis. The sensitivity analysis includes the following steps: Set different window lengths; Calculate the root mean square error between the denoised data and the reference data with different window lengths; Based on the trend of the root mean square error, the window length corresponding to when the root mean square error tends to stabilize and is close to its minimum value is selected as the optimal length.

[0028] Report timestamps based on sensor-sensed information For sequential, serial arrangement , To improve the spatiotemporal coherence of data tracks and construct a high-density spatiotemporal dataset. .

[0029] (4) In the formula, To combine time series data sets A function sorted chronologically. Given the high precision of the TDOA system, when... and When data points have the same timestamp, prioritize them. Data. Expand Right now: (5) In the formula, for The Middle The position vector of each measurement data point , , It provides its position information in the E, N, and U directions.

[0030] Subsequently, the nonlinear trajectories of the data in each direction are locally linearized using a sliding spatiotemporal window, and instantaneous velocity constraints are applied to identify abnormal trajectories. First, the instantaneous velocity vector of the UAV is calculated based on the time difference and position information between adjacent data points.

[0031] (6) In the formula, , as well as For data points The velocity vectors in the E, N, and U directions, respectively. For simplicity, we define... Identifiers for each dimension, , , as well as All are simplified as .

[0032] Then, according to the reported timestamp Sequentially, a sliding window is set for the velocity vector information of each dimension of the data. In each iteration, the box plot (BP) algorithm is executed to identify trajectories that violate motion laws based on anomalies in velocity vector information. The BP algorithm execution flow first calculates... The numerical quartiles of velocity vector information in each dimension ( , ) and interquartile range ( ).

[0033] (7) In the formula, express A collection of velocity vector information within the system. To calculate variables of The numerical calculation function. Then, using... , Corresponding Further calculations of the theoretical numerical upper bounds of velocity vectors in each dimension. and numerical lower bound .

[0034] (8) Finally, based on theoretical numerical limits, velocity vector data in various dimensions were used... The numerical situation defines the location information of data points in each dimension. Precision discrimination label (9) In the formula, when The values ​​are within the theoretical limits, indicating that the drone... The motion in this dimension at any given time conforms to the laws of physics, and will Assigned value The accuracy is high for time components; conversely, for velocity components exceeding the theoretical limit... , The value assigned is This means that the data component is an abnormal data component.

[0035] Abnormal measurement data screening stage.

[0036] Precision Label Abnormal data will be replaced with null values. It does not participate in the next round of BP algorithm iteration. This data update process can be seen in equation (10). (10) In the formula, for The state after each iteration, i.e., the denoised data points. After all data iterations are completed, regarding... Data sets It can be represented as (11) In the formula, for The state after each iteration. Based on this, according to the data source. Denoising data point set It can be separated into radar data sets. Collection of data with TDOA system .

[0037] S4. Perform a first-stage interpolation repair on the data gaps caused by filtering out the abnormal data components; the first-stage interpolation repair method adopts the linear time proportional interpolation method. The linear time-proportional interpolation method is as follows: For any data gap in an orthogonal direction, obtain the two adjacent data points that have undergone spatial decoupling before and after the time point. Based on the time and location information of two spatially decoupled data points, interpolated data is generated to fill the data gaps through linear scaling.

[0038] S5. Divide the overall data stream after the first stage of interpolation repair into multiple time windows, evaluate the continuity of the data stream in each time window based on the data density, and perform the second stage of interpolation repair on the data stream in the time window that is evaluated as having weak continuity, so as to enhance the spatiotemporal coherence of the overall data stream.

[0039] The second stage of interpolation is performed on the weakly continuous window, specifically as follows: Set a data density threshold; For a given time window, calculate the density of data points within it; If the data point density is lower than the data density threshold, the window is determined to be a weakly continuous window, and the midpoint of the time domain of the window is taken as the time reference, and interpolation is performed based on the data of the surrounding windows. If the data point density is not lower than the data density threshold, then the window is determined to be a strongly continuous window, and its original data is retained.

[0040] The UAV data processing method based on kinematic constraints also includes the following steps: By combining the data source information and precision labels of the data points, a binary confidence label is constructed for each data point after the interpolation repair is completed in the first and second stages; wherein, the binary confidence label is used to characterize the credibility of the data point for decision-making in subsequent data processing stages.

[0041] The method for constructing binary confidence labels is as follows: If all dimensional components of a data point are directly measured by a sensor and are all marked as high precision, then it is determined to be a high-confidence data point. If any dimension component of a data point originates from the interpolation generation in the second stage, or if any dimension component is marked as low precision, it is determined to be a low-confidence data point.

[0042] like Figure 3 As shown, to address the data loss issue caused by the removal of outlier measurement data and to provide a continuous data foundation for the spatiotemporal alignment of subsequent heterogeneous data, this embodiment provides a staged interpolation strategy, the specific of which is as follows: First, based on the temporal position of the outlier data, high-precision data is selected as the interpolation reference sample. Interpolation is performed on these positions, and the interpolated data is used to interpolate the null values ​​updated based on equation (10). To repair the flight path, the overall data stream is then divided into continuous time windows. Continuity is determined based on the data density within each window. For empty data windows with poor continuity, the data interpolated in the first step is used as samples to fill in the gaps, thereby enhancing the overall continuity of the data stream.

[0043] To fully utilize the high-precision data from all dimensions, data processing during the repair phase remains within a three-dimensional decoupled space. For simplicity, heterogeneous measurement datasets are used. and All can be unified as .right Each data point in the dataset is orthogonally decoupled, and the decoupled dataset can be represented as: (12) In the formula, , , They are respectively The components are located in the E, N, and U dimensions. Based on this, the precision labels corresponding to the data points in each dimension are also considered. , It can be further classified, and its classification expression is as follows: (13) In the formula, For a high-precision dataset of each dimension, the corresponding data for all its elements. The values ​​are all .on the contrary, The dataset consists of a set of outlier data for each dimension, which has been removed after being updated using equation (10). Subsequently, considering interpolation accuracy, the following steps are taken: As a reference sample, for Interpolation is performed on the time position to repair the data, thereby solving the problem of data damage caused by noise reduction.

[0044] Specifically, assuming in timestamp Data points in various dimensions Update results in the first interpolation stage The following can be calculated (for simplicity, Omit ): (14) In the formula, For timestamps For the interpolation function at a given location, considering computational efficiency, the interpolation method in the first interpolation stage is a linear time proportional interpolation method. Specifically, assuming that at... The data point belongs to the set First, obtain adjacent high-precision data points. , ,in , For the timestamps corresponding to the two points, then based on and The interpolated data are calculated using equation (15). The expression is as follows: (15) In the second interpolation stage, for the repaired datasets of each dimension... First, the overall data flow is divided into A continuous uniform time window Time window It is expressed as follows: (16) In the formula, Number the windows; The length of a single window is defined. Subsequently, within each window, a continuity criterion based on data density is set, and interpolation decisions are made according to the strength of continuity. For each window, a midpoint in time is defined. The data set update strategy in the window is as follows: (17) In the formula, This provides the updated window data for each dimension. If the method determines that an empty window is a weakly continuous window, then the midpoint of the window's time domain is used as the time reference for interpolation; conversely, if a non-empty window is determined to be a strongly continuous window, then the original data in the window is retained and the update is skipped, thereby minimizing the systematic error caused by the interpolated data.

[0045] To aid subsequent data processing, the data source and accuracy were considered. Perform a confidence level assessment. Set the data source as a prerequisite. Corresponding data source tags It can be represented as follows (18) like If all dimensional data sources are data obtained from sensor measurements, then determine their corresponding... for Conversely, if the data source for at least one dimension is interpolation, then it is determined that... for Simultaneously, construct three-dimensional data points Precision label The expression is as follows: (19) If all dimensional components of the data point have high-precision labels, then Assigned value If at least one dimension component corresponds to a low-precision label, then Assigned value The second interpolation data is set to have no corresponding precision labels for each dimension. If the data points have no corresponding precision labels for each dimension, this type of data only assists in spatiotemporal alignment. Assigned value Based on this, to evaluate the confidence level of each data point in the final dataset, the data source and precision labels are combined to construct a binary label (...). , ).

[0046] Specifically, since there is a coupling relationship between data source and data accuracy, impossible combinations are first eliminated. , ), ( , ), ( , Subsequently, measurements were taken using sensors. ) and short-term interpolation ( () is used as a high-confidence criterion to distinguish The logical expression for confidence level is: (20) In the formula, when for hour, If the sensor measurement is accurate, then it means... These are high-confidence data points; Indicates the logical NOT operation. The AND operation represents a logical AND operation; conversely, the AND operation represents a low-confidence data point. To process Indicator functions introduced by the value: (twenty one) Equation (20) can be further equivalent to: (twenty two); Therefore, data points with second interpolation data are classified as low-confidence data.

[0047] In other embodiments, experimental design and result verification are performed on the above-mentioned UAV data processing method based on kinematic constraints. The specific methods are as follows: Design of multi-altitude flight experiments for data collection in urban scenarios.

[0048] The experimental environment and flight experiment design in this embodiment primarily utilize data collected from actual measurements by the radar system and TDOA positioning system within the test range. The experimental area and sensing equipment details are as follows: Figure 4 As shown.

[0049] The equipment used in this experiment consisted of 20 TDOA base stations and one radar system. The deployment locations of the TDOA base stations and radar system in the test field, as well as the flight path of the UAV, are as follows: Figure 5 and Figure 6 As shown in the figure. The effective sensing range of the TDOA base station is approximately 800m, and the sensing range of the radar system is approximately 5000m. During the experiment, the UAV performed a flight mission along an arc-shaped test route. The RTK data results for the flight altitude were approximately 125m, 382m, and 482m, and the real-time altitude change trend is shown in the figure. Figure 7 , Figure 8 as well as Figure 9 As shown. To fully verify the effectiveness of the data fusion framework, the test flight path included typical UAV flight attitudes such as turning, climbing, and hovering.

[0050] Using the coordinates (119.986°, 30.284°, 0m) as the reference origin, the geographical coordinates of the test route were uniformly transformed to east, north, and sky coordinates. Flight experiments were conducted at three typical altitude levels: 120m, 400m, and 500m. The drone model used was DJI Mavic 4 Pro. The experimental configuration information for each altitude level is shown in Table 1.

[0051] Table 1 Experimental configuration information for each altitude level Flight test data collection results To eliminate invalid data and irrelevant variables and construct a high-quality dataset, this paper associates the target's trajectory with the "target ID" identifier of all measurement data, thereby extracting valid perception and positioning data containing spatial three-dimensional coordinates and reporting timestamp information. Statistical information of valid samples in each scenario is shown in Table 2.

[0052] Table 2 Statistical Analysis of Valid Measurement Data As shown in Table 2, the data length comparison reveals that, across all scenarios, the TDOA system's measurement data volume is significantly higher than that of radar measurement data, with a difference of approximately one order of magnitude. On one hand, the TDOA system employs a higher-frequency sampling mechanism, resulting in a higher data refresh rate. On the other hand, due to the obstruction and multipath effects of urban building clusters, the TDOA system is prone to generating a large number of discrete false location points during the calculation process. These abnormal location points increase the data volume but also lead to a decrease in SNR. This validates the necessity of outlier detection for the coarse measurement data of the TDOA system.

[0053] Furthermore, for heterogeneous measurement data, this section orthogonally decouples it, decomposing it into three directions: E, N, and U. Its spatial information description can be found in [the relevant section]. Figure 10 .

[0054] Due to its high sampling rate, the TDOA system has a high density of measurement data points, resulting in a relatively smooth trajectory. However, the data distribution exhibits certain clustering characteristics, which may contain a large number of false location points as mentioned earlier. These false data points will be identified and removed in subsequent processing. Furthermore, due to limitations in geometric configuration and signal obstruction, the TDOA system's measurement data may experience the interruption phenomenon shown in the figure when the quality of the target's radiated signal deteriorates. This indicates that in complex interference environments, the TDOA system struggles to achieve continuous localization and tracking of non-cooperative targets.

[0055] Compared to the TDOA system measurement data, the radar measurement data is sparser but more evenly distributed, and the track constructed from the data shows no significant loss, demonstrating the radar's strong continuous target tracking capability. However, the track constructed from its data exhibits some fluctuation, and the amplitude is greater than that of the track constructed from the TDOA system measurement data. Especially in the vertical direction, the track shows a "sawtooth" fluctuation, which may be caused by low-altitude multipath effects.

[0056] It is evident that the volatility of radar-measured flight tracks is sensitive to flight altitude. As the flight altitude increases from 120m to 500m, the amplitude of track volatility gradually increases, but exhibits a non-linear trend. At 400m, the track shows continuous small-range fluctuations, while at 500m, localized data shows significant fluctuations. This may be due to increased clutter interference in urban environments at higher elevation angles, leading to SNR attenuation of the echo signal and thus causing radar positioning errors.

[0057] Design of heterogeneous data processing solutions: This article is based on experimental settings. The duration is 5.0 seconds. Meanwhile, to quantitatively evaluate the performance of the phased spatiotemporal alignment scheme presented in this paper, a data matching success rate is introduced. Average synchronization time difference and the standard deviation of the synchronization time difference A comprehensive evaluation is conducted using three statistical indicators. The numerical calculation methods for the three indicators are as follows: (twenty three) (twenty four) (25) In the formula, For statistics A statistical function for the number of elements; For measuring data length using low sampling rate sensors; For matched data pairs The time difference in synchronization of heterogeneous data. This reflects the degree to which the algorithm successfully aligns multi-source heterogeneous data in the space-time, and the utilization rate of low-sampling-rate sensor measurement data. Used to quantify the synchronization deviation of matched heterogeneous data on the time axis. Additionally... By characterizing the discreteness of synchronization deviation, the robustness of the matching algorithm is revealed.

[0058] Track reconstruction results: Firstly, based on the linkage data of three flight altitude scenarios , and carried out research on The experiment investigated the sensitivity of window length to the denoising performance of the data reconstruction mechanism. The experiment aimed to explore the relationship between the model's real-time processing capability and its ability to identify outliers, and to determine the optimal window length to provide a data foundation for subsequent steps.

[0059] The experiment was set up with the original measurement data as the baseline control group, i.e., the window length was set to 0, and the data were processed through different lengths. Data after noise reduction by corresponding mechanism This is the experimental group. RMSE (Root Mean Square Error) was also introduced as a quantitative evaluation index. The experimental results are shown below. Figure 11 , Figure 12 , Figure 13 .

[0060] Figure 11 , Figure 12 , Figure 13 Demonstrates the noise reduction window under different test scenarios. The impact of window length on the performance of the abnormal track identification module in the data reconstruction mechanism. As the window length increases, the error evolution curves of each dimension of each sub-graph tend to stabilize, indicating that the algorithm's ability to detect abnormal data is approaching its upper bound. In 400m and 500m scenarios, such as Figure 11 and Figure 13 As shown, the curve exhibits distinct minimum points in the overall RMSE, occurring at window lengths of 12 and 6. Particularly at a window length of 6, the RMSE curve for the 400m scenario shows a significant precipitous drop, indicating that 6 is a critical performance inflection point, at which point the algorithm can effectively filter out most outlier data. As the window length exceeds 6, the curve tends to plateau.

[0061] Considering that further increasing the length would lead to a decrease in the overall efficiency of algorithm performance and real-time performance, this embodiment of the invention determines that 6 is... The optimal length is shown in Table 3. The denoising performance of the algorithm at this value is shown in Table 3.

[0062] Table 3. Outlier Detection Performance in Multiple Scenarios under Optimal Window Table 3 summarizes the positioning error of the original sensing and positioning data and the positioning accuracy before and after outlier removal under the optimal window scale. As shown in the table, the measurement error of the radar data is significantly higher than that of the TDOA system measurement data in all test scenarios. This is because active radar is more susceptible to interference from complex urban background clutter during UAV sensing and positioning. In contrast, the heterogeneous measurement data error of TDOA technology reaches its peak in the 400m positioning scenario under clock synchronization and effective multi-base station networking conditions, with the RMSE of the TDOA system data reaching 26.08m and the RMSE of the radar data reaching 53.82m. This may be because clutter interference and multipath effects are more severe in this scenario, causing a significant decrease in echo SNR, thus limiting sensor positioning. However, when the flight altitude increases to 500m, the UAV gradually escapes the building obstruction layer, and the main signal propagation path changes to LOS propagation. This effectively suppresses multipath effects, improves echo quality, and consequently leads to a significant decrease in the measurement error of the heterogeneous sensors.

[0063] It is worth noting that in the 500m scenario, the measurement error of the TDOA system was only 8.59m. On the one hand, the information transmission link was not severely obstructed by non-line-of-sight (NLOS) at this altitude; on the other hand, this may be because the UAV was in steady-state cruise mode, lacking high-dynamic flight maneuvers, which helped the system to continuously locate it, thereby reducing errors.

[0064] Furthermore, the outlier filtering mechanism proposed in this paper has achieved significant results for 400m scene data with large errors. Compared to Denoising data The positioning accuracy was improved by 48.42%. This indicates that under severe noise interference, the mechanism can effectively suppress large fluctuations and preserve the true flight path. Furthermore, observing the accuracy changes before and after data processing in a 120m scenario shows that… Compared to accuracy A slight decrease (5.88%) was observed. This reveals that while the mechanism effectively handles large fluctuations, it still has limitations in identifying the flight paths of drones with small-scale flight maneuvers. Specifically, it may misclassify some genuine abrupt changes as noise and remove them, leading to system errors. This indicates that the outlier identification and removal mechanism only performs preliminary data filtering. Therefore, it is necessary to further introduce algorithms to achieve deeper data processing. Based on the denoised data... Further implement a phased interpolation strategy to construct The reconstructed flight path is visible. Figure 14 .

[0065] Figure 14This describes the spatial distribution characteristics of the data during the staged interpolation process. As shown in the first row, the original linked data... The data contained a large amount of anomalous data caused by environmental interference or algorithm errors. Particularly in the 400m and 500m scenarios, some anomalous data exhibited drastic jumps in the vertical direction (U-axis). Combined with the normal data marked in red, it can be seen that the anomalous data identification program in the data reconstruction mechanism successfully captured and filtered out these outliers that violated the drone's motion patterns, thus achieving signal-to-noise separation.

[0066] like Figure 14 As shown in the second row, a staged interpolation method was used to repair the denoised positions on the damaged data after removing outliers. Compared with the data in the first row, the repaired data not only maintains the continuity of the track in the horizontal direction (EN plane), but also effectively suppresses large non-physical fluctuations in the U-axis direction, significantly improving the accuracy of the measured track.

[0067] However, the track still contains numerous interruptions due to invalid sensor observations, which hinders the spatiotemporal alignment of subsequent heterogeneous measurement data. Therefore, this invention further utilizes a second-stage interpolation process to conditionally interpolate the repaired heterogeneous data; the interpolated track is shown in the third and fourth rows. While interpolation significantly improves the overall continuity of the track for TDOA system measurement data, it still has limitations in large-scale temporal interruption regions. For example, when processing 400m scene data, the original TDOA system measurement data contains numerous temporal interruption regions. If the UAV's high-dynamic flight maneuvers, such as turning and climbing, occur within these regions, the numerical interpolation strategy cannot capture the nonlinear changes in the track. This leads to the algorithm constructing abnormal tracks that deviate from actual maneuvers, introducing systematic positioning errors. Furthermore, the reconstructed radar measurement data shown in the fourth row still exhibits significant fluctuations. This indicates that the data reconstruction mechanism can effectively suppress large fluctuations, but it struggles to handle high-frequency, small-range fluctuations, making smoothing impossible solely through numerical interpolation.

[0068] Compared with the prior art, the embodiments of the present invention have the following significant advantages: High-precision anomaly identification: By setting a sliding window and using the velocity vector information of each dimension of the data as kinematic constraint indicators, the box plot algorithm is executed with the velocity vector as a parameter in each iteration, thereby combining statistical methods and physical constraints to effectively identify outliers; it effectively distinguishes between anomalies caused by environmental noise and the actual maneuvering flight of UAVs, improving the accuracy of outlier identification.

[0069] Anti-coupling interference: The orthogonal decoupling strategy in three-dimensional space is adopted, which enables the data in each dimension to be processed independently, avoids the mutual influence between abnormal information in different dimensions, and preserves the effective observation information in each dimension to the maximum extent.

[0070] Ensuring spatiotemporal continuity: An innovative two-stage interpolation repair strategy is proposed. This strategy not only fills the data interruptions caused by denoising, but also solves the inherent discontinuity problem in the acquisition of raw sensor data, providing a more complete and coherent high-quality data stream for subsequent data processing (such as data fusion and target tracking).

[0071] Multimodal fusion friendly: The data output by this method is not only denoised and repaired, but also given confidence information for each data point. This provides an important basis for subsequent spatiotemporal alignment and fusion decisions of heterogeneous sensor data, and improves the intelligence level and effectiveness of the overall processing framework.

[0072] Wide adaptability: By locally linearizing the nonlinear trajectory, this method is applicable to non-cooperative UAV flight scenarios with high dynamics and significant nonlinear characteristics.

[0073] In other embodiments, a kinematically constrained unmanned aerial vehicle (UAV) data processing system is also provided, including: A data acquisition module is used to receive raw data from non-cooperative UAV perception and positioning from at least two heterogeneous sensors; The data denoising module is used to perform coordinate transformation and spatial decoupling on the original data to obtain a set of data components in multiple orthogonal directions; the set of data components is linearized based on a preset sliding spatiotemporal window, and instantaneous velocity is applied as kinematic constraint in each direction to identify and filter out abnormal data components that violate the kinematic constraint. The data interpolation and repair module is used to perform a first-stage interpolation and repair on the data gaps formed by filtering out the abnormal data components; the overall data stream after the first-stage interpolation and repair is divided into multiple time windows, the continuity of the data stream in each time window is evaluated based on the data density, and a second-stage interpolation and repair is performed on the data stream in the time window that is evaluated as having weak continuity, so as to enhance the spatiotemporal coherence of the overall data stream.

[0074] In some other embodiments, a storage medium is also provided, which stores a computer program or computer instructions that, when executed by a computer's processor, implement the steps of the above-described UAV data processing method based on kinematic constraints.

[0075] The storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart memory card, SD card, flash memory card, etc., mounted on the device. Furthermore, the storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0076] In other embodiments, a computer is also provided, including a memory and one or more processors, wherein executable code is stored in the memory, and when the one or more processors execute the executable code, the steps of the above-described UAV data processing method based on kinematic constraints are implemented.

[0077] The memory can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or RAM. The memory can also be an external storage device of any data processing device, such as a plug-in hard disk, smart memory card, SD card, flash memory card, etc., mounted on the device. Furthermore, the memory can include both internal storage units and external storage devices of any data processing device. The memory is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the concept and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A UAV data processing method based on kinematic constraints, characterized in that, Includes the following steps: Receive raw data from non-cooperative UAV perception and localization from at least two heterogeneous sensors; The original data is subjected to coordinate transformation and spatial decoupling to obtain a set of data components in multiple orthogonal directions; The data component set is linearized based on a preset sliding spatiotemporal window, and instantaneous velocities are applied in each direction as kinematic constraints to identify and filter out abnormal data components that violate the kinematic constraints. The first stage of interpolation repair is performed on the data gaps caused by filtering out the abnormal data components; The overall data stream after the first stage of interpolation repair is divided into multiple time windows. The continuity of the data stream in each time window is evaluated based on the data density. The data stream in the time window that is evaluated as having weak continuity is subjected to the second stage of interpolation repair to enhance the spatiotemporal coherence of the overall data stream.

2. The UAV data processing method based on kinematic constraints according to claim 1, characterized in that, The original data is subjected to coordinate transformation and spatial decoupling to obtain a set of data components in multiple orthogonal directions, specifically including the following steps: Obtain the original data points of the original data; wherein, the original data points include geographic coordinates and timestamps; Select a preset reference origin, and transform the geographic coordinates into positional components in the east, north, and sky directions under the station center coordinate system through a coordinate transformation matrix; The transformed location components are sorted chronologically based on the timestamps to form an ordered set of data components.

3. The UAV data processing method based on kinematic constraints according to claim 2, characterized in that, Identifying and filtering out anomalous data components that violate the kinematic constraints specifically includes the following steps: Within the sliding spacetime window, for each orthogonal direction, calculate the instantaneous velocity sequence of the data component in that direction; For the instantaneous velocity sequence in each direction, the box plot algorithm is executed to determine the theoretical normal range of instantaneous velocity values ​​in that direction; If the instantaneous velocity of a data point in any direction exceeds the theoretical normal range for that direction, then the component of the data point in that direction is determined to violate the kinematic constraints, and a low-precision label is assigned to that component. Data components with low-precision labels are filtered out as null values, and the steps of this claim are iteratively executed until all data has been processed.

4. The UAV data processing method based on kinematic constraints according to claim 3, characterized in that, It also includes the following steps: By combining the data source information and precision labels of the data points, a binary confidence label is constructed for each data point after the interpolation repair is completed in the first and second stages; wherein, the binary confidence label is used to characterize the credibility of the data point for decision-making in subsequent data processing stages.

5. The UAV data processing method based on kinematic constraints according to claim 4, characterized in that, The method for constructing binary confidence labels is as follows: If all dimensional components of a data point are directly measured by a sensor and are all marked as high precision, then it is determined to be a high-confidence data point. If any dimension component of a data point originates from the interpolation generation in the second stage, or if any dimension component is marked as low precision, it is determined to be a low-confidence data point.

6. The UAV data processing method based on kinematic constraints according to claim 1, characterized in that, The first stage of interpolation repair uses linear time-proportional interpolation. The linear time-proportional interpolation method is as follows: For any data gap in an orthogonal direction, obtain the two adjacent data points that have undergone spatial decoupling before and after the time point. Based on the time and location information of two spatially decoupled data points, interpolated data is generated to fill the data gaps through linear scaling.

7. The UAV data processing method based on kinematic constraints according to claim 1, characterized in that, The second stage of interpolation is performed on the weakly continuous window, specifically as follows: Set a data density threshold; For a given time window, calculate the density of data points within it; If the data point density is lower than the data density threshold, the window is determined to be a weakly continuous window, and the midpoint of the time domain of the window is taken as the time reference, and interpolation is performed based on the data of the surrounding windows. If the data point density is not lower than the data density threshold, then the window is determined to be a strongly continuous window, and its original data is retained.

8. The UAV data processing method based on kinematic constraints according to claim 1, characterized in that, The length of the sliding spatiotemporal window is predetermined to be optimal through sensitivity analysis. The sensitivity analysis includes the following steps: Set different window lengths; Calculate the root mean square error between the denoised data and the reference data for different window lengths; Based on the trend of the root mean square error, the window length corresponding to when the root mean square error tends to stabilize and is close to its minimum value is selected as the optimal length.

9. A system employing the UAV data processing method based on kinematic constraints as described in any one of claims 1 to 8, characterized in that, include: A data acquisition module is used to receive raw data from non-cooperative UAV perception and positioning from at least two heterogeneous sensors; The data denoising module is used to perform coordinate transformation and spatial decoupling on the original data to obtain a set of data components in multiple orthogonal directions; the set of data components is linearized based on a preset sliding spatiotemporal window, and instantaneous velocity is applied as kinematic constraint in each direction to identify and filter out abnormal data components that violate the kinematic constraint. The data interpolation and repair module is used to perform a first-stage interpolation and repair on the data gaps formed by filtering out the abnormal data components; the overall data stream after the first-stage interpolation and repair is divided into multiple time windows, the continuity of the data stream in each time window is evaluated based on the data density, and a second-stage interpolation and repair is performed on the data stream in the time window that is evaluated as having weak continuity, so as to enhance the spatiotemporal coherence of the overall data stream.

10. A computer, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they implement the steps of the UAV data processing method based on kinematic constraints as described in any one of claims 1 to 8.

Images