Method and device with improved radar resolution

1. A processor-implemented method, the method comprising:

executing computer readable code, stored in a non-transitory computer-readable storage medium, by a processor and configuring the processor, through the execution, to perform operations of:

identifying information of at least one of dimensions defining an original raw radar data;

generating a high-resolution training ground truth and a low-resolution training input from obtained original raw radar data based on the identified information; and

iteratively training a machine learning radar resolution increase model based on the low-resolution training input and the high-resolution training ground truth.

2. The method of claim 1 , wherein the dimensions defining the original raw radar data comprise one of a Doppler velocity, a horizontal angle, an elevation angle, and a range, or a combination of two or more of the Doppler velocity, the horizontal angle, the elevation angle, and the range.

3. The method of claim 1 , wherein the generating of the high-resolution training ground truth and the low-resolution training input comprises:

selecting low-resolution radar data from the original raw radar data; and

generating the low-resolution training input by preprocessing the selected low-resolution radar data.

4. The method of claim 3 , wherein the selecting of the low-resolution radar data comprises:

selecting, as the low-resolution radar data, radar data corresponding to a first part of antenna channels from the original raw radar data in an angular dimension.

5. The method of claim 3 , wherein the selecting of the low-resolution radar data comprises:

selecting, as the low-resolution radar data, radar data corresponding to a portion of chirp signals implemented for single scanning from the original raw radar data in a Doppler velocity dimension.

6. The method of claim 3 , wherein the selecting of the low-resolution radar data comprises:

selecting, as the low-resolution radar data, radar data corresponding to a first part of bandwidths allocated to the original raw radar data in a range dimension.

7. The method of claim 3 , wherein the generating of the low-resolution training input by preprocessing the selected low-resolution radar data comprises:

performing a range fast Fourier transform (FFT) process and a digital beamforming (DBF) process on the selected low-resolution radar data.

8. The method of claim 7 , wherein the generating of the low-resolution training input by preprocessing the selected low-resolution radar data comprises:

generating at least one of a range-Doppler map and a range-angle map from the low-resolution radar data.

9. The method of claim 1 , wherein the generating of the high-resolution training ground truth and the low-resolution training input comprises:

selecting high-resolution radar data from the original raw radar data; and

generating the high-resolution training ground truth by preprocessing the selected high-resolution radar data.

10. The method of claim 1 , wherein the radar resolution increase model comprises a neural network, the neural network comprising a convolution layer,

wherein the training of the radar resolution increase model comprises:

further training the radar resolution increase model based on direction-of-arrival (DoA) information estimated based on subspace type-based estimation.

11. The method of claim 1 , wherein the training the radar resolution increase model comprises training a temporary radar resolution increase model that comprises a generator model and a discriminator model,

wherein the training of the temporary radar resolution increase model comprises:

generating a temporary high-resolution output from the low-resolution training input based on the generator model;

training the generator model and the discriminator model based on a result output by the discriminator model from the temporary high-resolution output, the high-resolution training ground truth, and DoA information estimated based on subspace type-based estimation; and

in response to the training of the temporary radar resolution increase model being determined completed, removing the discriminator model from the temporary radar resolution increase model to generate the trained radar resolution increase model.

12. A processor-implemented radar data processing method comprising:

generating, by a radar sensor, raw radar data based on transmission of a radar signal by the sensor and reception of a radar signal by the sensor;

generating high-resolution output data from input data generated by preprocessing the generated raw radar data based on a trained radar resolution increase model; and

generating driving-related information based on the generated high-resolution output data.

13. The method of claim 12 , wherein the generating of the high-resolution output data comprises:

generating the input data by performing a range fast Fourier transform (FFT) process and a digital beamforming (DBF) process on the raw radar data.

14. The method of claim 12 , wherein the generating of the high-resolution output data comprises:

generating, as the input data, one of a range-Doppler map and a range-angle map; and

outputting, as the high-resolution output data, a map of a same type as a type of the input data from a map generated as the input data based on the resolution increase model.

15. The method of claim 12 , wherein the radar sensor is implemented in a mobile terminal,

wherein the generating of the raw radar data comprises:

generating, by the radar sensor, the raw radar data by transmitting a radar signal to at least a portion of an object and receiving a radar signal reflected from the portion of the object, and

the method further comprising:

recognizing at least one of a gesture associated with the object or an identity of the object based on the high-resolution output data, and indicating a result of the recognizing.

16. The method of claim 12 , wherein the radar sensor is implemented as a plurality of radar sensors in a vehicle,

the method further comprising:

transferring low-resolution raw radar data from the plurality of radar sensors to a central processor; and

generating, by the central processor, the high-resolution output data from the low-resolution raw radar data based on the radar resolution increase model.

17. A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of claim 12 .

18. A training device comprising:

a memory configured to store a resolution increase model; and

a processor configured to receive original raw radar data, and generate a high-resolution training ground truth and a low-resolution training input from the original raw radar data based on information corresponding to at least one of dimensions defining the original raw radar data, and train the resolution increase model such that the high-resolution training ground truth is output from the low-resolution training input.

19. A processor-implemented training method comprising:

transmitting, by a sensor, a radar signal to an object;

receiving a reflection of the transmitted radar signal from the object;

generating raw radar data from the received reflection of the transmitted radar signal;

perform a high-resolution preprocessing operation on the raw radar data to generate a high-resolution training ground truth data;

performing a low-resolution preprocessing operation on low-resolution radar data selected from the raw radar data to generate a low-resolution training input; and

training a generator model and a discriminator model based on the low-resolution training input, the high-resolution training ground truth, and direction-of-arrival (DoA) information estimated based on subspace type-based estimation.

20. The method of claim 19 , wherein the high-resolution preprocessing operation comprising one or more of a range fast Fourier transform (FFT) process, a digital beamforming (DBF) process, and a non-coherent integration process.

21. The method of claim 1 , further comprising providing the radar resolution increase model for detecting at least one object based on the radar resolution increase model.

22. The method of claim 1 , wherein generating comprises generating, as the low-resolution training input, a radar scan image from a low-resolution radar data of the original raw radar data, the radar scan image being one of a range-Doppler map and a range-angle map.