Unmanned aerial vehicle positioning method, training method for unmanned aerial vehicle positioning model and system using the same

Abstract

This disclosure provides a method for UAV localization, a training method for UAV localization models, and a corresponding system. The method includes: receiving a radio frequency signal stream from a target UAV through multiple antennas at each base station; inputting the radio frequency signal stream into a trained neural network model corresponding to each base station, so that the multiple trained neural network models output multiple prediction ranges, wherein each prediction range includes multiple predicted location points, and each predicted location point has corresponding coordinates and a confidence value; establishing a spatial grid corresponding to the localization system and the target UAV; mapping each predicted location point of the multiple prediction ranges to the spatial grid; accumulating the mapped confidence values ​​to obtain a probability weight value; and selecting the target grid point with the highest probability weight value as the predicted location of the target UAV.

Claims

1. A drone positioning method, applied to a positioning system comprising multiple base stations, each base station having multiple antennas and a corresponding trained neural network model, the method comprising: The radio frequency signal stream of a target drone is received through the multiple antennas of each base station; The radio frequency signal stream is input into the trained neural network model corresponding to each base station, so that the multiple trained neural network models output multiple estimated ranges, wherein each estimated range includes multiple estimated location points output by the corresponding trained neural network model, and each estimated location point has a corresponding coordinate and a corresponding confidence value; the multiple estimated ranges are obtained from the multiple base stations, and a spatial grid corresponding to the positioning system and the target UAV is established to map each estimated location point of the multiple estimated ranges to the spatial grid; for each grid point of the spatial grid, the confidence value of each estimated location point mapped to the grid point is accumulated to obtain a probability weight value of the grid point; And select the target grid point with the highest probability weight value as the estimated location of the target UAV.

2. The UAV positioning method as described in claim 1, wherein the step of accumulating the confidence value of each of the estimated location points mapped to the grid point includes: For each grid point in the spatial grid, the confidence values ​​of all the predicted location points mapped to that grid point are summed to obtain the probability weight value of that grid point.

3. The UAV positioning method as described in claim 1, wherein the step of selecting the target grid point with the highest probability weight value includes: When there are multiple grid points with the same highest probability weight value in the spatial grid, calculate the geometric centroid of the region formed by these multiple grid points, and use the geometric centroid as the estimated location.

4. The UAV positioning method as described in claim 1, wherein the spatial grid is a three-dimensional spatial grid, and each grid point of the three-dimensional spatial grid corresponds to a set of longitude, latitude and altitude coordinates.

5. The drone localization method as described in claim 1, wherein the confidence value reflects the degree of certainty of the trained neural network model that the location is the location of the target drone.

6. The UAV positioning method as described in claim 1, wherein the confidence value of each estimated location point within each estimated range is normalized so that the sum of all confidence values ​​within each estimated range is a predetermined value.

7. The UAV positioning method as described in claim 1, wherein the method is repeatedly executed over multiple consecutive time intervals and the probability weight values ​​obtained in each time interval are superimposed to form a time-series cumulative probability weight value, wherein the estimated position of the target UAV is determined based on the time-series cumulative probability weight value.

8. The drone positioning method as described in claim 1, wherein the target drone is a non-cooperative target, and the positioning system performs positioning by passively receiving the radio frequency signal stream broadcast by the target drone, without relying on the location information returned by the target drone.

9. A training method for a UAV positioning model, applied to a positioning system comprising multiple base stations, each base station having multiple antennas, the method comprising: A sampling drone is moved within a predetermined airspace, wherein the sampling drone broadcasts a radio frequency signal and transmits its spatial position during the movement; the radio frequency signal and the spatial position are received by each base station, and an elevation angle of the sampling drone relative to each base station is calculated in real time; when the elevation angle is greater than a predetermined threshold, each base station collects training data at a first sampling frequency; when the elevation angle is less than or equal to the predetermined threshold, each base station collects the training data at a second sampling frequency, wherein the first sampling frequency is greater than the second sampling frequency, wherein each piece of training data collected by each base station includes the radio frequency signal stream received by the multiple antennas of the base station at a sampling time point and the spatial position of the sampling drone at the sampling time point; and a neural network model corresponding to each base station is trained using the training data, so that each neural network model learns the correspondence between the radio frequency signal stream and the spatial position.

10. The training method as described in claim 9, wherein the spatial location includes three components: longitude, latitude, and altitude, which serve as label data for training the neural network model.

11. The training method as described in claim 9, wherein the neural network model is a convolutional neural network, the convolutional neural network including at least one convolutional layer for extracting waveform features of the radio frequency signal stream, at least one pooling layer for dimensionality reduction, and at least one fully connected layer for mapping feature vectors to the spatial location.

12. The training method as described in claim 9, wherein the method further comprises: During the movement of the sampling drone, each base station monitors the signal-to-noise ratio of the received radio frequency signal stream in real time. When the signal-to-noise ratio is lower than a predetermined quality threshold, the sampling frequency of the corresponding area is increased.

13. The training method as described in claim 9, wherein the method further comprises: During the operation of the positioning system, when the positioning accuracy is lower than a predetermined accuracy threshold, the sampling drone is controlled to collect supplementary data for the airspace area where the accuracy has decreased, and the neural network model is incrementally learned using the supplementary training data.

14. The training method as described in claim 9, wherein the method further comprises: Each base station monitors the training data collected relative to a spatial coverage distribution of the base station, the spatial coverage distribution including an azimuth coverage range and an elevation coverage range; when the spatial coverage distribution does not meet a predetermined coverage condition, a sparse area of ​​training data is identified; and based on the sparse area of ​​training data, the flight path of the sampling UAV is adjusted to supplement the training data corresponding to the sparse area of ​​training data, and the sampling UAV is controlled to fly within the predetermined airspace based on the adjusted flight path.

15. A drone positioning system, comprising: Multiple base stations, each with multiple antennas, are used to receive radio frequency signal streams from a target drone. The system also includes a server communicatively connected to the plurality of base stations. The server comprises a processor, a storage device, and a communication circuit unit. The storage device stores a plurality of trained neural network models and a plurality of program modules. Each trained neural network model corresponds to one of the plurality of base stations. The processor executes the plurality of program modules to perform the following operations: receiving radio frequency signal streams from each of the base stations via the communication circuit unit; inputting the radio frequency signal streams into the trained neural network model corresponding to each base station, so that the plurality of trained neural network models output a plurality of prediction ranges, wherein each prediction range includes a plurality of predicted location points output by the corresponding trained neural network model, and each predicted location point has a corresponding coordinate and a corresponding confidence value; establishing a spatial grid corresponding to the positioning system and the target drone, mapping each predicted location point of the plurality of prediction ranges to the spatial grid; and accumulating the confidence values ​​of each predicted location point mapped to the grid point for each grid point in the spatial grid to obtain a probability weight value for the grid point. And select the target grid point with the highest probability weight value as the estimated location of the target UAV.

16. A training system for a UAV positioning model, comprising: Multiple base stations, each with multiple antennas, are used to receive radio frequency signal streams from a sampling drone; The system also includes a server communicatively connected to the plurality of base stations. The server includes a processor, a storage device, and a communication circuit unit. The storage device stores a plurality of program modules. The processor executes the plurality of program modules to perform the following operations: controlling the sampling drone to move within a predetermined airspace via the communication circuit unit, wherein the sampling drone broadcasts the radio frequency signal stream and transmits back a spatial position of the sampling drone during movement; receiving the radio frequency signal stream and the spatial position via each base station through the communication circuit unit, and calculating the elevation angle of the sampling drone relative to each base station in real time; when the elevation angle is greater than a predetermined threshold, collecting training data at a first sampling frequency. When the elevation angle is less than or equal to the predetermined threshold, the training data is collected at a second sampling frequency, wherein the first sampling frequency is greater than the second sampling frequency, wherein each piece of training data collected by each base station includes the radio frequency signal stream received by the multiple antennas of the base station at a sampling time point and the spatial position of the sampling drone at the sampling time point; and a neural network model corresponding to each base station is trained with the training data, so that each neural network model learns the correspondence between the radio frequency signal stream and the spatial position, and the trained neural network models are stored in the storage device.

Images