Beidou grid space-time intelligent low-altitude sensing monitoring target data fusion method

Abstract

The application discloses a Beidou grid space-time intelligent low-altitude sensing target data fusion method, and particularly relates to the technical field of low-altitude sensing and edge computing data fusion, which comprises the following steps: acquiring trajectory records reported by multiple low-altitude sensing devices, projecting the trajectory records into grid sampling points through Beidou grid coding, generating motion codes according to the displacement of adjacent grid sampling points, and dividing fine trajectory sequences and coarse trajectory sequences according to the difference between the positioning error quantization value and the device reference error; for the fine trajectory sequences, taking the displacement of adjacent fine trajectory sampling points as motion vectors, taking the vertical direction of the fine trajectory sampling points as a rotation axis, and taking the inverse tangent value of the positioning error quantization value divided by the Beidou grid distance as a rotation angle; extrapolating the grid center through the fine trajectory sequences to form a local calculation boundary, projecting the coarse trajectory sampling points into a grid ball to generate projection residuals, and then distributing edge computing resources according to the projection residuals and performing trajectory fusion.

Description

Technical Field

[0001] This invention relates to the field of low-altitude sensing and monitoring and edge computing data fusion technology, and more specifically, to a method for fusion of target data for low-altitude sensing and monitoring based on BeiDou grid spatiotemporal intelligent. Background Technology

[0002] Currently, low-altitude target monitoring mainly uses radar, cameras, ADS-B, UAV detectors and other equipment to collect target trajectories, and converts latitude, longitude, altitude, timestamp and motion status into a unified Beidou grid spatiotemporal code. Then, the grid granularity is adjusted according to the positioning accuracy of the equipment and the airspace control level for matching and fusion between different equipment trajectories. In scenarios where urban logistics drone routes are close to airport airspace, high-precision positioning trajectories, radar ranging trajectories, and camera recognition trajectories will all converge into the same edge computing node. The node needs to complete candidate trajectory selection, similarity calculation, and target trajectory continuation within a millisecond time slice, while being limited by processor computing power, inference queue length, and cache capacity. In this scenario, if the grid is expanded according to the error range of low-precision equipment, different targets flying adjacently in the same short time will be included in the same candidate set, which manifests as a rapid increase in the number of candidate trajectory pairs, an extended queuing time for fusion calculation, and an increase in erroneous merging when adjacent targets intersect. If the grid is tightened according to the error range of high-precision equipment, the trajectory reported after camera obstruction, radar ranging offset, or spectrum detection delay will be difficult to enter the same target candidate range. This manifests as the same flight target being split into multiple target identifiers and the interruption of cross-device trajectory continuity. The reason is that the existing grid granularity adjustment is still based on the equipment type or regional level, and it is impossible to use high-precision trajectory to limit the calculation boundary of low-precision trajectory in a local spatiotemporal range. The technical problem to be solved by this application is: how to use high-precision trajectories to form a local fine-grid computing range when the accuracy of multi-source low-altitude trajectories is inconsistent and edge computing resources are limited, so that low-precision trajectories can participate in fusion computing only within the necessary range, thereby avoiding the expansion of the overall candidate scale by low-precision data and reducing the breakage of the same target across devices. Summary of the Invention

[0003] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a data fusion method for target sensing and monitoring based on BeiDou grid spatiotemporal intelligent. This method forms a local computational boundary by extrapolating the fine track sequence to the grid center, projects the coarse track sampling position into the grid sphere to generate projection residuals, allocates edge computing resources based on the projection residuals, and performs trajectory fusion to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for fusing spatiotemporal intelligent low-altitude sensing and monitoring target data based on the BeiDou grid, comprising: S1. Obtain the trajectory records reported by the multi-source low-altitude sensing device, project the trajectory records into grid sampling positions through Beidou grid coding, generate motion codes according to the displacement of adjacent grid sampling positions, and divide the fine track sequence and coarse track sequence by the difference sign after subtracting the equipment reference error from the positioning error quantization value. S2. For the precision track sequence, take the displacement of adjacent precision track sampling positions as the motion vector, take the vertical direction of the precision track sampling position as the rotation axis, take the arctangent value of the positioning error quantization value divided by the Beidou grid distance as the rotation angle, obtain the borrowing grid direction vector through the Rodriguez rotation transformation algorithm, and extrapolate the borrowing grid center of the next time slice along the borrowing grid direction vector. S3. Based on the coarse track sequence and the borrowed grid center, a borrowed grid sphere is generated with the borrowed grid center as the sphere center and the Beidou grid distance corresponding to the quantized value of the fine track positioning error as the radius. The Euclidean sphere projection algorithm is performed on the coarse track sampling position. When they coincide, the borrowed grid center is taken as the coarse track sampling position after projection. When they do not coincide, the direction from the borrowed grid center to the coarse track sampling position is normalized, multiplied by the radius, and added back to the borrowed grid center to obtain the coarse track sampling position after projection and the projection residual. Candidate trajectory pairs are output. S4. Read the remaining execution cycle of the current time slice of the central processing unit at the edge computing node, sort the candidate trajectory pairs in ascending order of projection residual, and deduct the execution cycle one by one from the first one. Write the candidate trajectory pairs that have been deducted into the execution quota and cache quota of the central processing unit, and write the coarse track sampling bits that have not been deducted into the coarse track summary, and output the resource table. S5. Read the candidate trajectory pairs from the resource table and perform fusion technology processing. Align the fine track sampling position and the projected coarse track sampling position to the same time slice. Generate fusion weights by reciprocal normalization of the positioning error quantization value. Calculate the fusion sampling positions and connect them to the same target trajectory chain. Output the fusion trajectory of the low-altitude sensing monitoring target.

[0005] In a preferred embodiment, S1 includes: S11. Obtain the trajectory records and the remaining execution cycle of the current time slice of the central processing unit. Establish the encoding sequence according to the arrival order of the trajectory records. Starting from the first position of the encoding sequence, deduct one execution cycle for each trajectory record read. Write the trajectory records that still have execution cycles after deduction into the record to be encoded. Output the record to be encoded and the remaining trajectory records. The time slice refers to the time unit obtained by dividing the low-altitude sensing time axis by the edge computing node according to a fixed sampling period. The integer quotient of the sampling time of each trajectory record divided by the time slice length is used as the time slice number, which is used to make the trajectory records reported by different low-altitude sensing devices in the same time unit enter the same encoding calculation batch. S12. Perform BeiDou grid coding on the record to be coded. First, project the position of the record to be coded onto the BeiDou plane grid to obtain the grid number. Then, rewrite the grid number as the time slice grid number by dividing the sampling time by the time slice length. Finally, fill in the vertical position of the time slice grid number with the integer quotient of the flight altitude divided by the layer distance and output the grid sampling position. S13. Read the adjacent grid sampling bits, subtract the previous grid sampling bit from the next grid sampling bit to get the grid displacement, and use the grid bits to remove the motion code based on the time slice difference; subtract the device reference error from the positioning error quantization value to get the error difference value, write the grid sampling bits with non-positive error difference value into the fine track sequence, write the grid sampling bits with positive error difference value into the coarse track sequence, and output the fine track sequence and the coarse track sequence. The positioning error quantization value refers to the positioning error radius given by the low-altitude sensing device corresponding to the trajectory record in this report, or the positioning error radius calculated by the edge computing node according to the device calibration error, reporting delay and the quality of the current sampling signal, and then converted into the number of grids according to the BeiDou grid distance. It is used to represent the uncertain range of the grid sampling position in the BeiDou grid. The equipment reference error refers to the reference positioning error written into the edge computing node after the low-altitude sensing equipment has completed calibration. The reference positioning error is converted into the number of grids according to the Beidou grid distance and used as the equipment reference error. It is used to subtract the positioning error quantization value and determine whether the grid sampling position is written into the fine track sequence or the coarse track sequence.

[0006] In a preferred embodiment, S2 includes: S21. Read the precision track sequence and the remaining execution cycle of the current time slice of the central processing unit. Generate adjacent precision track pairs according to the time slice order of the precision track sampling bits. Deduct one execution cycle for each adjacent precision track pair read. Write the adjacent precision track pairs that have been deducted into the rotation calculation queue. Write the precision track sampling bits that have not been deducted into the precision track summary. Output the rotation calculation queue and the precision track summary. S22. For adjacent precision orbit pairs in the rotation calculation queue, the motion vector is obtained by subtracting the previous precision orbit sampling bit from the subsequent precision orbit sampling bit. The vertical direction of the previous precision orbit sampling bit is used as the rotation axis, and the arctangent value of the positioning error quantization value divided by the Beidou grid distance is used as the rotation angle. The positive rotation vector is calculated by the Rodriguez rotation transformation algorithm, and the positive rotation vector is substituted back into the back-substitute motion vector by the opposite rotation angle. The positive rotation vector and the back-substitute difference are output. The BeiDou grid distance refers to the actual spatial interval between the centers of two adjacent plane grid numbers in the current airspace level. Edge computing nodes use the BeiDou grid distance to convert the quantized value of positioning error into the in-grid deflection scale, and use it as a unified spatial reference for rotation angle calculation and rounding. The process of calculating the positive rotation vector is as follows: First, divide the rotation axis by the length of the rotation axis to obtain the unit rotation axis. Then, calculate the cross product, dot product, and original vector components of the motion vector along the unit rotation axis. Subsequently, according to the Rodrigues rotation transformation algorithm, multiply the original vector components by the cosine of the rotation angle, multiply the cross product by the sine of the rotation angle, multiply the unit rotation axis by the product of the dot product and the product of the first minus the cosine of the rotation angle, and then add the three terms together to obtain the positive rotation vector.

[0007] In a preferred embodiment, S2 further includes: S23. Perform a consistency check on the forward rotation vector based on the back-substitution difference. When the back-substitution motion vector equals the motion vector after being rounded up according to the BeiDou grid distance, write the forward rotation vector into the valid rotation vector. When the back-substitution motion vector does not equal the motion vector after being rounded up according to the BeiDou grid distance, bisect the rotation angle and re-execute the Rodriguez rotation transformation algorithm until the back-substitution difference is rounded up to zero or the central processing unit's execution cycle is reduced, and output the valid rotation vector. S24. Based on the effective rotation vector, extrapolate the candidate center of the next time slice, connect the candidate center of the borrowed grid to the previous fine track sampling position to obtain the core vector, and use the grid position difference between the core vector and the motion vector as the center cost term. Write the candidate center of the borrowed grid corresponding to the first position of the center cost term in ascending order into the borrowed grid center, and output the borrowed grid center bound to the fine track sequence. The process of extrapolating the candidate center of the next time slice is as follows: read the effective rotation vector and the next precision orbit sampling position, divide the effective rotation vector by the time slice difference of the adjacent precision orbit sampling positions to obtain the single time slice rotation displacement, add the single time slice rotation displacement to the next precision orbit sampling position, and round it to the center of the grid number according to the Beidou grid distance to obtain the candidate center of the next time slice.

[0008] In a preferred embodiment, S3 includes: S31. Read the coarse track sequence, borrowing center and the remaining execution cycle of the current time slice of the central processing unit. Merge the coarse track sampling bits according to the same borrowing center to generate projection groups. Subtract the ball parameter calculation cycle once for each projection group. Write the projection group that has been subtracted into the projection calculation queue. Output the projection calculation queue and coarse track summary. The sphere parameter calculation cycle refers to the execution cycle consumed by the central processing unit to calculate the center and radius of the borrowed sphere for the same projection group and write it into the projection group cache. S32. For the projection group in the projection calculation queue, generate a borrowed grid sphere with the borrowed grid center as the sphere center and the Beidou grid distance corresponding to the quantized value of the precision orbit positioning error as the radius, and write the sphere center and radius of the borrowed grid sphere into the projection group cache, and output the projection group cache.

[0009] In a preferred embodiment, S3 further includes: S33. Read the coarse track sampling bit in the projection group buffer and the projection group. First, calculate the center distance from the coarse track sampling bit to the center of the borrowed grid. When the center distance is zero, write the borrowed grid center as the coarse track sampling bit after projection. When the center distance is not zero, divide the direction from the borrowed grid center to the coarse track sampling bit by the center distance to get the unit direction. Then multiply the unit direction by the radius and add it back to the borrowed grid center to get the coarse track sampling bit after projection. Use the distance from the coarse track sampling bit to the coarse track sampling bit after projection as the projection residual. S34. Generate candidate trajectory pairs based on the coarse track sampling position after projection, the projection residual, and the fine track sampling position bound to the borrowed grid center. Use the sum of the grid distance corresponding to the projection residual, the time slice difference, and the motion code difference as the candidate cost term. Take the first position of the candidate trajectory pairs corresponding to the same coarse track sampling position in ascending order of the candidate cost term and output the candidate trajectory pairs.

[0010] In a preferred embodiment, S4 includes: S41. Read the candidate trajectory pairs, projection residuals and the remaining execution cycle of the current time slice of the central processing unit. Divide the projection residuals by the Beidou grid distance to obtain the integer quotient as the residual bits. Generate the execution order with the residual bits as the first sorting key and the arrival order of the candidate trajectory pairs as the second sorting key and output it. S42. Read candidate trajectory pairs according to execution order, add the number of fine trajectory sampling bits in the candidate trajectory pair to the number of coarse trajectory sampling bits after projection to obtain the execution cost, and use the number of bytes written to the fusion calculation buffer of the candidate trajectory pair to obtain the cache cost, and output the execution cost and cache cost. The fusion computing buffer refers to the memory area in the edge computing node where candidate trajectory pairs are temporarily stored before entering the fusion computing process. The contents written to it include fine trajectory sampling bits, coarse trajectory sampling bits after projection, projection residuals, and resource table indexes. The central processing unit reads candidate trajectory pairs from the fusion computing buffer according to the resource table and performs subsequent fusion computing.

[0011] In a preferred embodiment, S4 further includes: S43. Starting from the first position of execution order, read the candidate trajectory pair, execution consumption and cache consumption. First, subtract the execution consumption from the remaining execution cycle of the current time slice of the central processing unit. Then, subtract the cache consumption from the current cache balance of the edge computing node. When both differences are non-negative, write the candidate trajectory pair into the execution quota and cache quota of the central processing unit. Stop deduction when at least one difference is negative and output the resource table. The buffer space refers to the number of writable memory bytes of the edge computing node that have not yet been occupied by the fusion computing buffer in the current time slice. Its value is obtained by subtracting the number of bytes already written to the fusion computing buffer from the total number of bytes allocated to the fusion computing buffer in the current time slice. It is used to determine whether the candidate trajectory pair can be written to the fusion computing buffer. S44. Read the remaining candidate trajectory pairs after stopping the deduction, bind the coarse track sampling bits in the remaining candidate trajectory pairs with the corresponding projection residuals and write them into the coarse track summary, bind the coarse track summary to the same time slice of the resource table, and output the resource table and the coarse track summary.

[0012] In a preferred embodiment, S5 includes: S51. Read the resource table and fusion calculation buffer, read the candidate trajectory pairs according to the execution order in the resource table, and extract the fine track sampling bits, the coarse track sampling bits after projection, the projection residual and the resource table index from the candidate trajectory pairs, and output the trajectory pairs to be fused. The process of outputting the trajectory pairs to be merged is as follows: Candidate trajectory pairs are read bit by bit from the fusion calculation buffer according to the execution order in the resource table. If the candidate trajectory pair read has both fine track sampling bits and projection coarse track sampling bits, the fine track sampling bits, projection coarse track sampling bits, projection residuals and resource table indexes are combined into the trajectory pairs to be merged. If the fine track sampling bits or projection coarse track sampling bits are missing, the candidate trajectory pair is skipped and the next candidate trajectory pair is read. S52. Read the time slice number of the fine track sampling position and the time slice number of the coarse track sampling position after projection in the trajectory to be fused. Multiply the motion code of the coarse track sampling position after projection by the time slice difference between the two to obtain the compensation displacement. Add the compensation displacement to the coarse track sampling position after projection and output the coarse track sampling position in the same slice.

[0013] In a preferred embodiment, S5 further includes: S53. Read the fine track sampling position, the coarse track sampling position in the same piece and the corresponding positioning error quantization value of the trajectory pair to be fused. Take the reciprocal of the positioning error quantization value respectively, and then divide each reciprocal by the sum of the two reciprocals to obtain the fine track weight and the coarse track weight. Output the fusion weight. S54. Calculate the weighted bits of the fine track sampling bits and the coarse track sampling bits in the same area according to the fusion weight. Add the two weighted bits together and round them according to the Beidou grid distance to obtain the fusion sampling bits. Then, connect the fusion sampling bits to the target trajectory chain according to the resource table index and output the fusion trajectory of the low-altitude sensing and monitoring target. The calculation process for the weighted bits is as follows: multiply the grid number, vertical bit, and time slice number of the fine track sampling bit by the fine track weight to obtain the fine track weighted bit; multiply the grid number, vertical bit, and time slice number of the coarse track sampling bit in the same slice by the coarse track weight to obtain the coarse track weighted bit; then add the fine track weighted bit and the coarse track weighted bit together to obtain the composite grid bit used to generate the fused sampling bit; A target trajectory chain refers to a trajectory data chain formed by writing the same low-altitude monitoring target into the fused sampling bits according to the time slice number, using the same low-altitude monitoring target as an index, to represent the positional succession relationship of the same low-altitude monitoring target in different time slices; The fused sampling position refers to the grid position obtained by calculating the fine track sampling position and the coarse track sampling position in the same time frame according to the fusion weight and rounding it according to the Beidou grid distance. The fused sampling position includes the planar grid position and the vertical position in the same time frame and serves as a continuation node in the target trajectory chain.

[0014] The technical effects and advantages of this invention are as follows: This scheme generates grid centers by using fine track sequences and restricts the sampling positions of coarse tracks to projection within the grid sphere. Low-precision trajectories only participate in fusion within the local boundaries of the fine track, which relatively alleviates the expansion of candidate scale and the breakage of target continuity. The remaining execution cycle of the current time slice is used to filter trajectory records before entering the Beidou grid coding, so that the edge computing nodes can prioritize processing the data that can be handled, thereby reducing the occupation of invalid codes and the queuing of subsequent calculations. Perform a Rodrigues rotation transformation on the fine orbit motion vector and extrapolate the grid center so that the motion direction and error scale participate in the boundary generation, which relatively improves the candidate continuation after occlusion reproduction and delayed reporting. In the Euclidean sphere projection algorithm, the sampling position of the coarse track after projection and the projection residual are generated, so that the deviation of the coarse track is converted into a sortable computational quantity, which relatively suppresses low-precision trajectories and expands the overall search range. The execution order is generated by the projection residual, and the execution cycle and cache space of the central processing unit are deducted simultaneously, so that the candidate trajectory pairs enter the fusion according to resource constraints, which relatively improves the edge-side fusion scheduling. Attached Figure Description

[0015] Figure 1 This is a technical roadmap for the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1 The present invention provides a method for fusing spatiotemporal intelligent low-altitude sensing and monitoring target data based on the BeiDou grid, comprising: S1. Obtain the trajectory records reported by the multi-source low-altitude sensing device, project the trajectory records into grid sampling positions through Beidou grid coding, generate motion codes according to the displacement of adjacent grid sampling positions, and divide the fine track sequence and coarse track sequence by the difference sign after subtracting the equipment reference error from the positioning error quantization value. Trajectory records from low-altitude monitoring sites are entered into edge computing nodes according to different devices, different levels of precision, and different arrival times. The central processing unit has a limited amount of coding computation that it can handle within the current time slice. Therefore, in this implementation, the remaining execution cycle of the current time slice is used to limit the trajectory records entering the encoding, the trajectory records are written into a uniform grid sampling bit, and finally, the fine track sequence and the coarse track sequence are generated according to the same error quantization caliber. The trajectory record is a recording unit capable of reconstructing the low-altitude target position, sampling time, flight altitude, equipment source, and positioning error source; the BeiDou plane grid origin, BeiDou grid distance, airspace level, and layer distance are loaded by the edge computing nodes in the current monitoring area, and the time slice length is registered by the current low-altitude monitoring task and does not change with a single trajectory record during operation; the implementation process includes the following steps: After the trajectory record reaches the edge computing node, it is first encoded in S11 according to the arrival order, so that the execution cycle of the central processing unit of the current time slice becomes the constraint of the encoding entry point. After the edge computing node reads the trajectory record, the arrival order of the trajectory record and the remaining execution cycle of the current time slice of the central processing unit, it reads the trajectory record one by one from the first position of the encoding order. Before reading, it determines whether the remaining execution cycle of the current time slice is positive. If it is positive, it deducts one execution cycle. When the value after deduction is not less than zero, the trajectory record is written into the record to be encoded. When the value after deduction is less than zero, the edge computing node rolls back the current deduction and writes the trajectory record read this time and the trajectory record not yet read in the encoding sequence into the remaining trajectory record; The time slices here are obtained by dividing the low-altitude sensing time axis by a fixed sampling period. The integer quotient of the sampling time divided by the time slice length is used as the time slice number. The same time slice number corresponds to the same code calculation batch. The original location, sampling time, and flight altitude to be encoded have different forms of expression in different low-altitude sensing devices. S12 rewrites the three into the same grid sampling bit through Beidou grid coding. After the edge computing node reads the record to be encoded, the origin of the BeiDou plane grid, the BeiDou grid distance, the time slice length and the layer spacing, it converts the position of the record to be encoded into a plane distance relative to the origin of the BeiDou plane grid, and obtains the grid number by dividing the plane distance by the BeiDou grid distance. The time slice number is then obtained by dividing the sampling time by the time slice length and quotienting the integer, and the time slice number is merged into the grid number to form the time slice grid number; Continue to obtain the vertical position by dividing the flight altitude by the layer spacing, and then fill the time frame number with the vertical position to obtain the grid sampling position; The remainders generated when rounding up the planar distance, sampling time, and flight altitude are not written into the grid sampling bits, but are reserved for the conversion of positioning error quantization values. They are written into the remaining trajectory records when the encoded record is missing position, sampling time, or flight altitude. After completing the writing of the grid sampling bits, S13 divides the fine track sequence and coarse track sequence according to the error difference formed by this report; Edge computing nodes read adjacent grid sampling bits according to device origin and time slice number. When multiple grid sampling bits exist in the same time slice number for the same device origin, the last grid sampling bit is taken according to the arrival order to participate in adjacent calculations. The edge computing node obtains the grid displacement by subtracting the previous grid sampling bit from the next grid sampling bit, and obtains the time slice difference by subtracting the time slice number of the previous grid sampling bit from the time slice number of the next grid sampling bit. When the time slice difference is not zero, use the cell to remove the motion code obtained from the time slice difference; when the time slice difference is zero, write an empty motion code. The positioning error radius is divided by the BeiDou grid distance to obtain the number of error grids. If there is a remainder in the number of error grids, it is rounded up to obtain the positioning error quantization value. The reference positioning error after equipment calibration is converted into the equipment reference error according to the same conversion method. The edge computing node then subtracts the equipment reference error from the positioning error quantization value to obtain the error difference value. The grid sampling bits with non-positive error difference values ​​are written into the fine track sequence, and the grid sampling bits with positive error difference values ​​are written into the coarse track sequence. The fine track sequence and the coarse track sequence are output. After the above processing is completed, the fine track sequence entering the subsequent S2 has a unified grid position, unified time slice number and unified error quantization caliber. The coarse track sequence entering the subsequent S3 also retains traceable motion codes and error differences. Track records that are not handled by the central processing unit or have missing fields are left in the remaining track records to wait for the next time slice processing. In practical applications: After radar records, camera records, and ADS-B records near the airport's airspace are collected and entered into the edge computing node, the edge computing node first filters out the records to be encoded according to the remaining execution cycle of the current time slice. Then, it writes the position of the record to be encoded as a grid number, the sampling time as a time slice number, and the flight altitude as a vertical position. Finally, it uses the error difference between the positioning error quantization value and the equipment reference error to complete the separation of the fine orbit sequence and the coarse orbit sequence, so that subsequent grid rotation and Euclidean sphere projection only read the data that has been uniformly encoded.

[0018] S2. For the precision track sequence, take the displacement of adjacent precision track sampling positions as the motion vector, take the vertical direction of the precision track sampling position as the rotation axis, take the arctangent value of the positioning error quantization value divided by the Beidou grid distance as the rotation angle, obtain the borrowing grid direction vector through the Rodriguez rotation transformation algorithm, and extrapolate the borrowing grid center of the next time slice along the borrowing grid direction vector. When the precision track sequence enters the borrowing calculation stage, the edge computing node needs to select the adjacent precision track pair that can complete the rotation calculation from the precision track sampling bits within the remaining execution cycle of the current time slice of the central processing unit, and use the Rodriguez rotation transformation algorithm to convert the precision track movement direction into an extrapolable borrowing direction. In this embodiment, the precision orbit sampling position is represented by a grid consisting of a plane grid number, a vertical position, and a time slice number. The BeiDou grid distance is used for distance conversion between the centers of the plane grid numbers, and the layer distance is used for distance conversion between the centers of the vertical positions. The rotation axis is taken as the unit direction in the current BeiDou grid coordinate system that is consistent with the direction of vertical position increment. This implementation process includes the following steps: The central processing unit first constrains the rotation calculation entry point based on the remaining execution cycle of the current time slice. S21 reads the precision track sampling bits in the precision track sequence according to the time slice order and forms adjacent precision track pairs by combining two adjacent precision track sampling bits. After the edge computing node reads the precision track sequence and the remaining execution cycle of the current time slice of the central processing unit, it arranges the precision track sampling bits in order of time slice number. When the time slice difference between adjacent precision track sampling bits is positive, adjacent precision track pairs are generated. When the time slice difference is zero, the next precision track sampling bit is written into the precision track summary. Before reading each adjacent precision track pair, first determine whether the remaining execution cycle of the current time slice of the central processing unit is sufficient to deduct one execution cycle. If the value after deduction is not less than zero, write the adjacent precision track pair into the rotation calculation queue. If the value after deduction is less than zero, roll back the current deduction and write the next precision track sampling bit and subsequent precision track sampling bit in the adjacent precision track pair into the precision track summary. The rotation calculation queue is read by S22, and the precision track summary retains the sampling bit index, time slice number, and reason for not participating in the rotation calculation, for processing in the next time slice; The Rodriguez rotation transformation is performed in S22 based on three-dimensional grid coordinates. After the edge computing node reads the adjacent precision orbit pair from the rotation computing queue, it subtracts the plane grid number and vertical position of the previous precision orbit sampling position from the plane grid number and vertical position of the subsequent precision orbit sampling position to obtain the motion vector. The time slice number is only used to calculate the time slice difference and does not participate in the vector coordinate subtraction. When the previous precision track sampling position lacks a vertical position or the rotation axis length is zero, the adjacent precision track pair is written into the precision track summary. When the rotation axis length is non-zero, the edge computing node divides the rotation axis by the rotation axis length to obtain the unit rotation axis, divides the actual distance corresponding to the positioning error quantization value by the Beidou grid distance to obtain the grid deflection scale, and then takes the arctangent value of the grid deflection scale to obtain the rotation angle. Then calculate the cross product, dot product, and original vector components of the motion vector and the unit rotation axis. Multiply the original vector components by the cosine of the rotation angle, multiply the cross product by the sine of the rotation angle, multiply the unit rotation axis by the dot product, and subtract the cosine of the rotation angle. Add the three terms together to obtain the positive rotation vector. The edge computing node then performs a Rodriguez back substitution on the forward rotation vector using the opposite rotation angle to obtain the back substitution motion vector, and outputs the forward rotation vector and the back substitution difference. The back substitution difference is the difference between the grid coordinates of the back substitution motion vector after rounding according to the BeiDou grid distance and the grid coordinates of the motion vector. After the rotation result enters S23, the grid position drift caused by excessive rotation angle needs to be eliminated by back substitution difference. The edge computing node reads the forward rotation vector, back substitution motion vector, motion vector and the remaining execution cycle of the current time slice of the central processing unit. The back substitution motion vector is converted into grid position coordinates according to the Beidou grid distance and layer distance. Then, it is subtracted from the grid position coordinates of the motion vector to obtain the rounded result of back substitution difference. When the rounding result of the back-substitution difference is zero, the edge computing node writes the positive rotation vector into the valid rotation vector; When the rounding result of the back substitution difference is not zero, the edge computing node bisects the rotation angle and re-executes the Rodriguez rotation transformation and reverse back substitution when there are enough remaining execution cycles to deduct one rotation iteration cycle. Each re-execution deducts one rotation iteration cycle. When the rounding result of the back iteration difference is zero, a valid rotation vector is output. If the remaining execution cycle is less than one rotation iteration cycle, the iteration stops and the current adjacent fine track pair is written into the fine track summary. Adjacent fine track pairs that have not generated a valid rotation vector do not enter S24. The borrowed grid center is generated by extrapolation of the effective rotation vector in S24, and the borrowed grid center that can be read by the coarse orbit projection is selected by the center cost term. The edge computing node reads the effective rotation vector, the next precision track sampling position, the previous precision track sampling position, and the time slice difference. It divides the effective rotation vector by the time slice difference to obtain the single time slice rotation displacement. Then, it adds the single time slice rotation displacement to the grid coordinates of the next precision track sampling position and rounds it to the grid center and vertical position center according to the Beidou grid distance and layer distance to obtain the borrowed grid candidate center for the next time slice. Edge computing nodes use the borrowed candidate center to connect the previous precision track sampling position to obtain the kernel vector, and use the sum of the absolute values ​​of the grid position difference between the kernel vector and the motion vector as the center cost term; When multiple candidate centers exist, the edge computing nodes select the first one in ascending order of the center cost item. If the center cost items are the same, they are selected in ascending order of the grid distance corresponding to the time slice difference. If they are still the same, they are selected in order of generation of candidate centers and the selection result is written into the borrow center. The borrow center bound to the precision track sequence is output. After processing by this implementation method, the fine track sequence no longer only provides the difference between adjacent positions, but also forms a borrowed grid center with back-substitution verification and resource deduction records. Subsequently, S3 can directly read the borrowed grid center to establish a borrowed grid sphere and limit the coarse track sampling position to the local calculation boundary of the fine track. In practical applications: In urban low-altitude routes, after a logistics drone forms a fine track sequence by a high-precision positioning device, the edge computing node first selects adjacent fine track pairs according to the remaining execution cycle of the current time slice, then uses the sampling bits of the previous and next fine tracks to form a motion vector, and generates a rotation angle according to the positioning error quantization value; The forward rotation vector is validated by reverse substitution to generate a valid rotation vector. The valid rotation vector is extrapolated to obtain the borrowed center of the next time frame, so that the coarse track sampling position generated by the subsequent radar or camera only participates in the projection calculation around this borrowed center.

[0019] S3. Based on the coarse track sequence and the borrowed grid center, a borrowed grid sphere is generated with the borrowed grid center as the sphere center and the Beidou grid distance corresponding to the quantized value of the fine track positioning error as the radius. The Euclidean sphere projection algorithm is performed on the coarse track sampling position. When they coincide, the borrowed grid center is taken as the coarse track sampling position after projection. When they do not coincide, the direction from the borrowed grid center to the coarse track sampling position is normalized, multiplied by the radius, and added back to the borrowed grid center to obtain the coarse track sampling position after projection and the projection residual. Candidate trajectory pairs are output. When the coarse track sequence enters the projection process, the edge computing node first uses the borrowed grid center obtained by S2 as the local computing boundary, and then merges the coarse track sampling positions around the same borrowed grid center into the same projection group. In this way, the central processing unit only needs to calculate the center and radius of the borrowed sphere once for a projection group. The coarse track sampling bits in the projection group share the same projection group buffer. Then, the Euclidean sphere projection algorithm is executed one by one, and the projection results are converted into candidate trajectory pairs. The borrowed sphere surface serves as the boundary for coarse orbit sampling points before participating in candidate fusion. The coarse orbit summary stores the coarse orbit sampling points that were not included in the current time-slice projection calculation, their corresponding borrowed center, and the reason for not processing. This implementation process includes the following steps: The resource consumption of projection calculation is first compressed at the sphere parameter reuse point. In S31, the edge computing node reads the coarse track sequence, the borrowed grid center and the remaining execution cycle of the current time slice of the central processing unit, and concatenates the plane grid number, vertical position and time slice number of the borrowed grid center into a merge key, and writes the coarse track sampling bits with the same merge key into the same projection group. Before each projection group is written into the projection calculation queue, the central processing unit deducts one sphere parameter calculation cycle. Projection groups with a deduction result not less than zero enter the projection calculation queue. If the deduction result is less than zero, the current deduction is rolled back, and the coarse track sampling bits in the projection group are written into the coarse track summary. The sphere parameter calculation cycle is the execution cycle consumed by the central processing unit to calculate the center and radius of the borrowed sphere and write it into the projection group cache. The projection calculation queue is read by S32. The coarse track summary records the coarse track sampling bit index, borrowed center, and the reason for incomplete sphere parameter calculation. The center and radius of the sphere are written in S32 one at a time according to the projection group to avoid repeated reading of the fine track error for each coarse track sampling bit in the projection group. The edge computing node reads the projection group from the projection computing queue, and reads the corresponding borrowed grid center and the quantized value of the precision track positioning error bound to the borrowed grid center; When the quantized value of the precision positioning error is expressed in terms of the number of grids, the radius is obtained by multiplying the quantized value of the precision positioning error by the BeiDou grid distance. When the quantized value of the precision positioning error is expressed in terms of distance, the radius is directly taken from the quantized value of the precision positioning error. Edge computing nodes generate borrowed spheres with the borrowed grid center as the sphere center and the radius as the sphere center, and write the sphere center and radius into the projection group cache; When the projection group lacks the borrowing center or the quantization value of the fine track positioning error, the coarse track sampling bits in the projection group are written into the coarse track summary, and the projection group cache is not generated. The projection of a single coarse rail sampling position is completed in S33. Before the calculation, the plane grid number and the vertical position are converted to the same three-dimensional space unit. The edge computing node reads the projection group cache and coarse track sampling bit, converts the plane grid number of the coarse track sampling bit into the plane grid center coordinates, converts the vertical position into the layer center coordinates according to the layer spacing, and then calculates the Euclidean distance from the coarse track sampling bit to the borrowed grid center as the center distance; when the center distance is zero, the borrowed grid center is written as the coarse track sampling bit after projection, and the projection residual is zero. When the center distance is not zero, the direction from the borrowed grid center to the coarse rail sampling position is divided by the center distance to obtain the unit direction. The unit direction is multiplied by the radius and added back to the borrowed grid center to obtain the projected coarse rail sampling position located on the surface of the borrowed grid sphere. The Euclidean distance from the coarse rail sampling position to the projected coarse rail sampling position is written as the projection residual. If the coarse rail sampling position lacks a plane grid number, vertical position, or time slice number, it is written into the coarse rail summary. In S34, the candidate trajectory pair is simultaneously constrained by the projection residual, the time slice relationship, and the motion code relationship. The edge computing node reads the coarse track sampling position after projection, the projection residual, the fine track sampling position bound to the grid center, the time slice number, and the motion code. It converts the projection residual into a distance value in the BeiDou grid unit, multiplies the time slice difference between the coarse track sampling position and the fine track sampling position after projection by the single time slice grid to obtain the grid distance corresponding to the time slice difference, and then multiplies the grid position difference of the motion codes of the two by the BeiDou grid to obtain the grid distance corresponding to the motion code difference. The candidate cost item is obtained by adding the grid spacing corresponding to the projection residual, the time slice difference, and the motion code difference. When multiple candidate trajectory pairs are generated for the same coarse track sampling position, the first one is written into the candidate trajectory pair in ascending order of the candidate cost item. When the candidate cost items are the same, the first one is written into the candidate trajectory pair in ascending order of the projection residual. If they are still the same, the first one is written into the candidate trajectory pair in the generation order. The candidate trajectory pair is then output. After the projection process is completed, the coarse track sampling position is restricted to the local calculation boundary formed by the fine track borrowing grid center. The projection group buffer reduces the calculation of repeated spherical parameters, and the projection residual provides a directly sortable calculation basis for the allocation of the central processing unit execution quota and buffer quota in S4. In practical applications: When the radar coarse track and the camera coarse track fall near the center of the same logistics drone fine track grid, the edge computing node first merges the coarse track sampling bits according to the grid center, then writes the sphere center and radius to the projection group once, and then projects each coarse track sampling bit onto the grid sphere surface, and selects candidate trajectory pairs according to the candidate cost terms, so that the coarse track data enters the subsequent resource allocation around the local boundary of the fine track.

[0020] S4. Read the remaining execution cycle of the current time slice of the central processing unit at the edge computing node, sort the candidate trajectory pairs in ascending order of projection residual, and deduct the execution cycle one by one from the first one. Write the candidate trajectory pairs that have been deducted into the execution quota and cache quota of the central processing unit, and write the coarse track sampling bits that have not been deducted into the coarse track summary, and output the resource table. After the candidate trajectory pairs enter the resource allocation stage, the edge computing nodes convert the projection residuals formed by S3 into sortable residual bits, and then combine the remaining execution cycles of the current time slice of the central processing unit and the cache balance of the fusion computing buffer to determine which candidate trajectory pairs enter the fusion computing. In this embodiment, the execution cycle is the basic computing quota that the central processing unit can deduct within the current time slice; the fusion computing buffer is a temporary memory area before the candidate trajectory pair enters S5; and the resource table is used to record the execution quota, cache quota, and buffer write position obtained by the candidate trajectory pair within the current time slice. This implementation process includes the following steps: The projection residual is first rewritten into residual bits suitable for sorting and reading by the central processing unit. After S41 reads the candidate trajectory pair, the projection residual and the remaining execution cycle of the current time slice of the central processing unit, it divides the projection residual by the Beidou grid distance to obtain the integer quotient as the residual bit, and retains the original value of the projection residual to enter the subsequent recording along with the candidate trajectory pair. Edge computing nodes arrange candidate trajectory pairs using the residual bit as the sorting key. When the residual bits are the same, they are arranged according to the arrival order of the candidate trajectory pairs to obtain the execution order. Candidate trajectory pairs with missing projection residuals are not included in the execution order generation, and the corresponding coarse track sampling bits are written into the coarse track summary, with the execution order available for S42 to read. After the execution order is determined, each candidate trajectory pair needs to be converted into the amount of resources that the central processing unit and the fusion computing buffer can deduct. S42 reads the candidate trajectory pair according to the execution order and uses the number of fine track sampling bits in the candidate trajectory pair plus the number of coarse track sampling bits after projection to obtain the execution cost. Each sampling bit corresponds to one fusion read and one grid bit write, and the execution cost is used as the deduction base for the central processing unit's execution cycle; The number of bytes required for edge computing nodes to recompute candidate trajectory pairs and write them into the fusion computing buffer is calculated by adding the fixed byte length of the fine trajectory sampling bits, the coarse trajectory sampling bits after projection, the projection residual, and the resource table index. When a candidate trajectory lacks fine track sampling bits or coarse track sampling bits after projection, execution cost and cache cost are not generated, and the corresponding coarse track sampling bits are written into the coarse track summary. Resource deduction proceeds in S43 in an execution order that cannot be skipped. The edge computing node reads the candidate trajectory pair, execution cost, and cache cost from the first position of the execution order. First, it subtracts the execution cost from the remaining execution cycle of the current time slice of the central processing unit, and then subtracts the cache cost from the current cache balance of the edge computing node. When both differences are non-negative, the edge computing node writes the candidate trajectory pair into the central processing unit's execution quota and cache quota, and writes the time slice number, execution order, candidate trajectory pair index, central processing unit execution quota, cache quota, and fusion computing buffer write position into the resource table; When at least one difference is negative, the edge computing node stops further deduction, allowing candidate trajectory pairs with higher residual bits to occupy the current time slice resources first. The buffer space is obtained by subtracting the number of bytes already written to the fusion compute buffer from the total number of bytes allocated to the fusion compute buffer in the current time slice; Candidate trajectory pairs for which the current time slice resource has not been obtained need to be retained in the subsequent processing chain. S44 reads the remaining candidate trajectory pairs after stopping the deduction, binds the coarse track sampling bits in the remaining candidate trajectory pairs with the corresponding projection residuals and writes them into the coarse track summary, and binds the coarse track summary to the same time slice in the resource table. The coarse track summary should at least save the coarse track sampling bit index, projection residual, corresponding borrowed grid center, candidate trajectory pair index, and the reason for not obtaining execution quota, so that it can re-participate in projection or resource allocation in the next time slice; The resource table and the coarse track summary are output together. The resource table is read by S5, and the coarse track summary is read for subsequent time slice recalculation. After processing by this implementation method, the projection residuals of candidate trajectory pairs are converted into execution order. The execution cycle of the central processing unit and the buffer space of the fusion calculation buffer jointly limit the fusion entry point. The resource table clearly records the candidate trajectory pairs that can enter S5, and the coarse track summary retains the coarse track sampling bits that have not entered the fusion calculation. In practical applications: After multiple radar coarse tracks and camera coarse tracks generate candidate trajectory pairs through S3, the edge computing nodes first generate residual bits according to the projection residuals and arrange them in the execution order. Then, they calculate the execution cost and cache cost one by one. The candidate trajectory pairs that the current time slice resources can handle are written into the resource table. The coarse track sampling bits after the resources are insufficient are entered into the coarse track summary to avoid low residual candidate trajectory pairs being squeezed out of the fusion calculation by later-arriving or high-occupancy candidate trajectory pairs.

[0021] S5. Read the candidate trajectory pairs from the resource table and perform fusion technology processing. Align the fine track sampling position and the projected coarse track sampling position to the same time slice. Generate the fusion weight by normalizing the reciprocal of the positioning error quantization value. Calculate the fusion sampling position and connect it to the same target trajectory chain. Output the fusion trajectory of the low-altitude sensing monitoring target. After the resource table completes the allocation of execution quota and cache quota for candidate trajectory pairs, the fusion processing stage only reads the candidate trajectory pairs that have obtained the current time slice resources, and fills in the coarse track sampling bits after projection to the time slice where the fine track sampling bits are located. Then, it forms the fusion weight according to the positioning error quantization value, and finally connects the fusion sampling bits to the target trajectory chain. In this embodiment, the resource table index points to the precision track sequence identifier bound to the candidate trajectory pair, the target trajectory chain uses the precision track sequence identifier as the index of the same low-altitude monitoring target, the motion code is the single time-slice grid displacement, and the fused sampling bit is the continuation node of the target trajectory chain in the corresponding time slice; the implementation process includes the following steps: The resource table has given the reading order of candidate trajectory pairs and the writing position of the fusion calculation buffer. S51 retrieves data from the fusion calculation buffer along this order to prevent candidate trajectory pairs that have not obtained execution quota from entering the fusion calculation. The edge computing node reads the resource table and the fusion computing buffer, locates the candidate trajectory pairs bit by bit according to the execution order in the resource table, and reads the fine track sampling bits, the coarse track sampling bits after projection, the projection residual and the resource table index from the candidate trajectory pairs; When the candidate trajectory pair read contains both fine trajectory sampling bits and projected coarse trajectory sampling bits, the edge computing node combines the fine trajectory sampling bits, projected coarse trajectory sampling bits, projection residuals, and resource table indexes into a trajectory pair to be fused. When a candidate trajectory pair lacks fine trajectory sampling bits or coarse trajectory sampling bits after projection, the edge computing node writes the candidate trajectory pair index, missing items and corresponding coarse trajectory sampling bits into the coarse trajectory summary, and continues to read the next candidate trajectory pair in the execution order, and outputs the trajectory pair to be fused. The time alignment process is completed by S52. The key point is to move the coarse track sampling position after projection to the time slot where the fine track sampling position is located, so as to ensure that the subsequent weight calculations are performed within the same time unit. The edge computing node reads the time slice number of the fine track sampling position and the time slice number of the coarse track sampling position after projection in the trajectory to be fused. The time slice difference is obtained by subtracting the time slice number of the coarse track sampling position from the time slice number of the fine track sampling position. When the time difference is positive, the coarse rail sampling position after projection is pushed forward by the corresponding number of single time-slice grid displacements along the motion code direction; when the time difference is negative, the coarse rail sampling position after projection is pushed back by the corresponding number of single time-slice grid displacements along the opposite direction of the motion code; when the time difference is zero, the grid position of the coarse rail sampling position after projection is retained. When the motion code is empty, the edge computing node writes the trajectory pair to be fused into the coarse track summary, but does not output the coarse track sampling bits of the same piece; When the motion code is valid, the edge computing node multiplies the motion code of the projected coarse rail sampling bit by the time slice difference to obtain the filler displacement, and adds the filler displacement to the projected coarse rail sampling bit, and outputs the coarse rail sampling bit of the same slice. In S53, the error quantization value is converted into fusion weight. The smaller the positioning error quantization value, the higher the proportion of the corresponding sampling bit in the fusion sampling bit. The edge computing node reads the fine track sampling bit, the coarse track sampling bit in the same piece of trajectory pair to be fused, and the corresponding positioning error quantization value. When the positioning error quantization value is zero, the error quantization value corresponding to one Beidou grid distance is used for the reciprocal calculation to avoid division by zero. The edge computing nodes calculate the reciprocal of the quantized value of the positioning error of the fine track sampling position and the reciprocal of the quantized value of the positioning error of the coarse track sampling position in the same piece, respectively. Then, the fine track weight is obtained by dividing the corresponding reciprocal of the fine track sampling position by the sum of the two reciprocals, and the coarse track weight is obtained by dividing the corresponding reciprocal of the coarse track sampling position in the same piece by the sum of the two reciprocals. When any positioning error quantization value is missing, the edge computing node writes the trajectory pair to be fused into the coarse track summary, records the reason for the missing positioning error quantization value, and outputs the fusion weight; The trajectory linking is completed in S54. The edge computing nodes calculate the fusion sampling bits according to the fusion weight and write the fusion sampling bits into the corresponding target trajectory chain through the resource table index. The edge computing node multiplies the planar grid position and vertical position of the fine track sampling position by the fine track weight to obtain the fine track weighted position. It also multiplies the planar grid position and vertical position of the coarse track sampling position in the same area by the coarse track weight to obtain the coarse track weighted position. Finally, the fine track weighted position and the coarse track weighted position are added together to obtain the composite grid position. Since S52 has already filled the coarse track sampling bits after projection to the time slot where the fine track sampling bits are located, the time slot number of the fused sampling bits is directly taken from the time slot number of the fine track sampling bits. The composite grid is rounded to the nearest Beidou grid distance to become a planar grid. If the remainder is half a Beidou grid distance, it is rounded up. If the remainder is less than half a Beidou grid distance, the integer quotient is retained. The same rounding rule is used for vertical grids according to the layer spacing. Edge computing nodes find the target trajectory chain corresponding to the precision track sequence identifier according to the resource table index, and connect the fused sampling bits to the target trajectory chain according to the time slice number, and output the fused trajectory of the low-altitude sensing monitoring target; After the fusion process is completed, the candidate trajectory pairs in the resource table are converted into fusion sampling bits in the target trajectory chain. The coarse track sampling bits after projection are filled in by motion code and enter the same time slice. The positioning error quantization value is controlled by reciprocal normalization to control the fusion weight. Records with missing motion code, missing sampling bits or missing positioning error quantization value are entered into the coarse track summary to wait for subsequent processing. In practical applications: After the fine orbit sampling bits of the same UAV target have occupied the current time-slice fusion resources, the edge computing node reads the corresponding candidate trajectory pair from the fusion computing buffer, adds the radar coarse orbit projection bits to the time slice where the fine orbit is located according to the motion code, calculates the fusion weight according to the positioning error quantization value of the two, and finally writes the fused sampling bits into the target trajectory chain corresponding to the UAV.

[0022] Among them Figure 1 It should be noted that: This figure illustrates the complete technical route of this method from multi-source low-altitude trajectory recording input to low-altitude sensing and monitoring target fusion trajectory output; First, S1 completes trajectory admission and BeiDou grid encoding under edge computing resource constraints, and separates the fine track sequence and coarse track sequence. The fine track sequence enters S2, and the next time-slice grid center is extrapolated through Rodrigues rotation transformation. The coarse track sequence and the grid center enter S3 together, and candidate trajectory pairs and projection residuals are generated through Euclidean sphere projection. Then, S4 reads the candidate trajectory pairs, projection residuals, and the execution cycle and cache balance of the central processing unit, generates a resource table, and writes the data of the unacquired resources into the coarse track summary. Finally, S5 performs in-slice filling, error reciprocal normalization, and trajectory linking according to the resource table, and outputs the fused trajectory of the low-altitude sensing and monitoring target. The solid arrows in the diagram represent the main processing flow and data transfer relationships, such as the sequential flow of fine track sequence, coarse track sequence, borrowed center, candidate track pair, projection residual, and resource table between each step; the dashed arrows represent the write-back relationship when resources are insufficient, that is, the coarse track data that has not obtained the execution resources of the current time slice in S4 is written into the coarse track summary and returned to the projection processing link corresponding to S3, waiting for the subsequent time slice to continue to participate in projection and resource allocation.

[0023] Working principle: This scheme first receives trajectory records reported by multi-source low-altitude sensing devices from edge computing nodes, and filters out records that can be encoded according to the remaining execution cycle of the central processing unit in the current time slice. Then, the position, time and altitude are uniformly written into grid sampling bits through Beidou grid encoding, and the fine track sequence and coarse track sequence are divided according to the difference between the quantized value of the positioning error and the reference error of the device. The fine track sequence is used to generate the local calculation boundary. The edge computing nodes obtain the motion vector based on the adjacent fine track sampling bits, and extrapolate the borrowed grid center of the next time slice through the Rodrigues rotation transformation algorithm. The coarse track sequence generates a borrowed grid sphere around the borrowed grid center, and obtains the projected coarse track sampling bits and projection residuals through the Euclidean sphere projection algorithm. The projection residuals are further converted into the execution order and cache writing basis of the central processing unit, and the candidate trajectory pairs of the resources are obtained for fusion calculation. Finally, fusion sampling bits are generated according to the alignment of the same time slice and the reciprocal normalization weight of the error, connected to the same target trajectory chain, and the fusion trajectory of the low-altitude sensing monitoring target is output. In airport airspace, urban logistics routes, or low-altitude inspection channels, radar, cameras, ADS-B, and UAV detection equipment simultaneously report target trajectories within the same airspace. Edge computing nodes first encode these trajectories into BeiDou grid sampling points and use the stable positioning trajectory as the fine track to establish the grid center for the next time slice. When radar ranging deviates or cameras are blocked and re-form a coarse track, the coarse track does not expand the entire candidate search range but is projected onto the boundary of the fine track grid sphere. The projection residual then determines whether to occupy the current central processing unit and cache resources. In this way, when multiple UAV targets flying adjacently intersect, are blocked, or are delayed in reporting, target continuation and trajectory fusion can still be completed with limited edge computing resources, reducing candidate expansion and erroneous merging caused by the coarse track.

[0024] The above description is merely 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 spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data fusion method for target sensing and monitoring based on BeiDou grid spatiotemporal intelligent technology, characterized in that: include: S1. Obtain the trajectory records reported by the multi-source low-altitude sensing device, project the trajectory records into grid sampling positions through Beidou grid coding, generate motion codes according to the displacement of adjacent grid sampling positions, and divide the fine track sequence and coarse track sequence by the difference sign after subtracting the equipment reference error from the positioning error quantization value. S2. For the precision track sequence, take the displacement of adjacent precision track sampling positions as the motion vector, take the vertical direction of the precision track sampling position as the rotation axis, take the arctangent value of the positioning error quantization value divided by the Beidou grid distance as the rotation angle, obtain the borrowing grid direction vector through the Rodriguez rotation transformation algorithm, and extrapolate the borrowing grid center of the next time slice along the borrowing grid direction vector. S3. Based on the coarse track sequence and the borrowed grid center, a borrowed grid sphere is generated with the borrowed grid center as the sphere center and the Beidou grid distance corresponding to the quantized value of the fine track positioning error as the radius. The Euclidean sphere projection algorithm is performed on the coarse track sampling position. When they coincide, the borrowed grid center is taken as the coarse track sampling position after projection. When they do not coincide, the direction from the borrowed grid center to the coarse track sampling position is normalized, multiplied by the radius, and added back to the borrowed grid center to obtain the coarse track sampling position after projection and the projection residual. Candidate trajectory pairs are output. S4. Read the remaining execution cycle of the current time slice of the central processing unit at the edge computing node, sort the candidate trajectory pairs in ascending order of projection residual, and deduct the execution cycle one by one from the first one. Write the candidate trajectory pairs that have been deducted into the execution quota and cache quota of the central processing unit, and write the coarse track sampling bits that have not been deducted into the coarse track summary, and output the resource table. S5. Read the candidate trajectory pairs from the resource table and perform fusion technology processing. Align the fine track sampling position and the projected coarse track sampling position to the same time slice. Generate fusion weights by reciprocal normalization of the positioning error quantization value. Calculate the fusion sampling positions and connect them to the same target trajectory chain. Output the fusion trajectory of the low-altitude sensing monitoring target.

2. The method for fusing target data based on BeiDou grid-based spatiotemporal intelligent low-altitude sensing and monitoring according to claim 1, characterized in that: S1 includes: S11. Obtain the trajectory records and the remaining execution cycle of the current time slice of the central processing unit. Establish the encoding sequence according to the arrival order of the trajectory records. Starting from the first position of the encoding sequence, deduct one execution cycle for each trajectory record read. Write the trajectory records that still have execution cycles after deduction into the record to be encoded. Output the record to be encoded and the remaining trajectory records. S12. Perform BeiDou grid coding on the record to be coded. First, project the position of the record to be coded onto the BeiDou plane grid to obtain the grid number. Then, rewrite the grid number as the time slice grid number by dividing the sampling time by the time slice length. Finally, fill in the vertical position of the time slice grid number with the integer quotient of the flight altitude divided by the layer distance and output the grid sampling position. S13. Read the adjacent grid sampling bits, subtract the previous grid sampling bit from the next grid sampling bit to obtain the grid displacement, and use the grid bits to remove the motion code based on the time slice difference; subtract the device reference error from the positioning error quantization value to obtain the error difference value, write the grid sampling bits with non-positive error difference value into the fine track sequence, write the grid sampling bits with positive error difference value into the coarse track sequence, and output the fine track sequence and the coarse track sequence.

3. The method for fusing target data based on BeiDou grid-based spatiotemporal intelligent low-altitude sensing and monitoring according to claim 2, characterized in that: S2 includes: S21. Read the precision track sequence and the remaining execution cycle of the current time slice of the central processing unit. Generate adjacent precision track pairs according to the time slice order of the precision track sampling bits. Deduct one execution cycle for each adjacent precision track pair read. Write the adjacent precision track pairs that have been deducted into the rotation calculation queue. Write the precision track sampling bits that have not been deducted into the precision track summary. Output the rotation calculation queue and the precision track summary. S22. For adjacent precision orbit pairs in the rotation calculation queue, the motion vector is obtained by subtracting the previous precision orbit sampling bit from the subsequent precision orbit sampling bit. The vertical direction of the previous precision orbit sampling bit is used as the rotation axis, and the arctangent value of the positioning error quantization value divided by the Beidou grid distance is used as the rotation angle. The positive rotation vector is calculated by the Rodriguez rotation transformation algorithm, and the positive rotation vector is substituted back into the back-substitute motion vector by the opposite rotation angle. The positive rotation vector and the back-substitute difference are output.

4. The method for fusing target data based on BeiDou grid-based spatiotemporal intelligent low-altitude sensing and monitoring according to claim 3, characterized in that: S2 also includes: S23. Perform a consistency check on the forward rotation vector based on the back-substitution difference. When the back-substitution motion vector equals the motion vector after being rounded up according to the BeiDou grid distance, write the forward rotation vector into the valid rotation vector. When the back-substitution motion vector does not equal the motion vector after being rounded up according to the BeiDou grid distance, bisect the rotation angle and re-execute the Rodriguez rotation transformation algorithm until the back-substitution difference is rounded up to zero or the central processing unit's execution cycle is reduced, and output the valid rotation vector. S24. Based on the effective rotation vector, extrapolate the candidate center of the next time slice, connect the candidate center of the borrowed grid to the previous fine track sampling position to obtain the core vector, and use the grid position difference between the core vector and the motion vector as the center cost term. Write the candidate center of the borrowed grid corresponding to the first position of the center cost term in ascending order into the borrowed grid center, and output the borrowed grid center.

5. The method for fusing target data based on BeiDou grid-based spatiotemporal intelligent low-altitude sensing and monitoring according to claim 4, characterized in that: S3 includes: S31. Read the coarse track sequence, borrowing center and the remaining execution cycle of the current time slice of the central processing unit. Merge the coarse track sampling bits according to the same borrowing center to generate projection groups. Subtract the ball parameter calculation cycle once for each projection group. Write the projection group that has been subtracted into the projection calculation queue. Output the projection calculation queue and coarse track summary. S32. For the projection group in the projection calculation queue, generate a borrowed grid sphere with the borrowed grid center as the sphere center and the Beidou grid distance corresponding to the quantized value of the precision orbit positioning error as the radius, and write the sphere center and radius of the borrowed grid sphere into the projection group cache, and output the projection group cache.

6. The data fusion method for target sensing and monitoring based on BeiDou grid spatiotemporal intelligent technology as described in claim 5, characterized in that: S3 also includes: S33. Read the coarse track sampling bit in the projection group buffer and the projection group. First, calculate the center distance from the coarse track sampling bit to the center of the borrowed grid. When the center distance is zero, write the borrowed grid center as the coarse track sampling bit after projection. When the center distance is not zero, divide the direction from the borrowed grid center to the coarse track sampling bit by the center distance to get the unit direction. Then multiply the unit direction by the radius and add it back to the borrowed grid center to get the coarse track sampling bit after projection. Use the distance from the coarse track sampling bit to the coarse track sampling bit after projection as the projection residual. S34. Generate candidate trajectory pairs based on the coarse track sampling position after projection, the projection residual, and the fine track sampling position bound to the borrowed grid center. Use the sum of the grid distance corresponding to the projection residual, the time slice difference, and the motion code difference as the candidate cost term. Take the first position of the candidate trajectory pairs corresponding to the same coarse track sampling position in ascending order of the candidate cost term and output the candidate trajectory pairs.

7. The method for fusing target data based on BeiDou grid-based spatiotemporal intelligent low-altitude sensing and monitoring according to claim 6, characterized in that: S4 includes: S41. Read the candidate trajectory pairs, projection residuals and the remaining execution cycle of the current time slice of the central processing unit. Divide the projection residuals by the Beidou grid distance to obtain the integer quotient as the residual bits. Generate the execution order with the residual bits as the first sorting key and the arrival order of the candidate trajectory pairs as the second sorting key and output it. S42. Read candidate trajectory pairs according to execution order, use the number of fine trajectory sampling bits in the candidate trajectory pair plus the number of coarse trajectory sampling bits after projection to obtain the execution cost, and use the number of bytes written to the fusion calculation buffer of the candidate trajectory pair to obtain the cache cost, and output the execution cost and cache cost.

8. The method for fusing target data based on BeiDou grid spatiotemporal intelligent low-altitude sensing and monitoring according to claim 7, characterized in that: S4 also includes: S43. Starting from the first position of execution order, read the candidate trajectory pair, execution consumption and cache consumption. First, subtract the execution consumption from the remaining execution cycle of the current time slice of the central processing unit. Then, subtract the cache consumption from the current cache balance of the edge computing node. When both differences are non-negative, write the candidate trajectory pair into the execution quota and cache quota of the central processing unit. Stop deduction when at least one difference is negative and output the resource table. S44. Read the remaining candidate trajectory pairs after stopping the deduction, bind the coarse track sampling bits in the remaining candidate trajectory pairs with the corresponding projection residuals and write them into the coarse track summary, bind the coarse track summary to the same time slice of the resource table, and output the resource table and the coarse track summary.

9. The method for fusing target data based on BeiDou grid-based spatiotemporal intelligent low-altitude sensing and monitoring according to claim 8, characterized in that: S5 includes: S51. Read the resource table and fusion calculation buffer, read the candidate trajectory pairs according to the execution order in the resource table, and extract the fine track sampling bits, the coarse track sampling bits after projection, the projection residual and the resource table index from the candidate trajectory pairs, and output the trajectory pairs to be fused. S52. Read the time slice number of the fine track sampling position and the time slice number of the coarse track sampling position after projection in the trajectory to be fused. Multiply the motion code of the coarse track sampling position after projection by the time slice difference between the two to obtain the compensation displacement. Add the compensation displacement to the coarse track sampling position after projection and output the coarse track sampling position in the same slice.

10. The method for fusing target data based on BeiDou grid-based spatiotemporal intelligent low-altitude sensing and monitoring according to claim 9, characterized in that: S5 also includes: S53. Read the fine track sampling position, the coarse track sampling position in the same piece and the corresponding positioning error quantization value of the trajectory pair to be fused. Take the reciprocal of the positioning error quantization value respectively, and then divide each reciprocal by the sum of the two reciprocals to obtain the fine track weight and the coarse track weight. Output the fusion weight. S54. Calculate the weighted bits of the fine track sampling bits and the coarse track sampling bits in the same area according to the fusion weight. Add the two weighted bits together and round them according to the Beidou grid distance to obtain the fused sampling bits. Then, connect the fused sampling bits to the target trajectory chain according to the resource table index and output the fused trajectory of the low-altitude sensing and monitoring target.

Images