Detection and classification of unmanned aerial vehicles
A method and system for real-time and automated detection and classification of aerial objects, such as e.g. UAVs, on a radar plot level from digital radar images, wherein the digital radar images are created from radar return signals obtained by a conventional radar such as e.g. a continuous wave radar (e.g. FMCW radar), a phased-array radar, or a pulse radar, as opposed to using specialized micro-Doppler radars for classifying UAVs based on the Doppler effect created by the rotors or propellers of the UAVs. Further, the method and system allow for tracking the UAVs based on the digital radar images.
1. A computer-implemented method for classifying an aerial object represented on a digital radar image, said aerial object being an aerial object that may pose a collision risk for air traffic, said method comprising:
receiving a chronological series of digital radar images originating from a radar system, said digital radar images each having a total area formed by individual pixels, each of said digital radar images comprising a radar plot of said aerial object,
combining radar plots of said aerial object into a sequence of radar plots of said aerial object,
obtaining from said sequence of radar plots of said aerial object a track for said aerial object,
for each of said digital radar images selecting a sub-area of said total area, said sub-area comprising said radar plot of said aerial object,
subjecting said sub-area to a deep learning model for determining whether said radar plot represents an aerial object belonging to one class of aerial objects, and
concluding that said track represents an aerial object belonging to said one class when it is determined for a plurality of said radar plots in said sequence that the radar plot subjected to said deep learning model belongs to said one class.
2. The method according claim 1 , comprising creating said digital radar image from a signal received from a radar system.
3. The method according to claim 1 , wherein said radar system comprises at least one of a phased-array radar, a pulse radar, and a FMCW radar.
4. The method according to claim 1 , wherein said step of selecting a sub-area of said total area comprises applying a detection algorithm to said digital radar image for detecting and locating in said digital radar images a radar plot of said aerial object.
5. The method according to claim 4 , wherein said step of applying said detection algorithm for detecting and locating said radar plot comprises filtering static clutter and dynamic clutter.
6. The method according to claim 1 , comprising applying at least one of a distance correction, a static clutter filtration, and a long-term average correction to said digital radar images, prior to applying said deep learning model.
7. The method according to claim 1 , wherein said sub-area comprises a lesser number of pixels compared to said total area.
8. The method according claim 1 , wherein said sub-area has a rectangular outline.
9. The method according to claim 1 , wherein said one class is a class of unmanned aerial vehicles (UAVs).
10. The method according to claim 1 , wherein said deep learning model is trained for classifying radar plots in sub-areas according to said one class.
11. The method according to claim 1 , wherein said deep learning model is a trained deep learning model.
12. The method according to claim 1 , wherein said deep learning model is a trained convolutional neural network.
13. The method according to claim 1 , wherein said step of selecting for each of said digital radar images a sub-area of said total area comprises cutting out said sub-area from said digital radar image and generating an image cutout comprising said sub-area.
14. Use of the method according to claim 1 to determine if an aerial object represented in a radar plot in a digital radar image is an unmanned aerial vehicle.
15. A computer system for classifying an aerial object represented on a digital radar image, said aerial object being an aerial object that may pose a collision risk for air traffic, said computer system comprising a deep learning model and a processor configured to:
receive a chronological series of digital radar images originating from a radar system, said chronological series of digital radar images each having a total area formed by individual pixels, each of said digital radar images comprising a radar plot of said aerial object,
combine radar plots of said aerial object into a sequence of radar plots of said aerial object,
obtain from said sequence of radar plots of said aerial object a track for said aerial object,
for each of said digital radar images select a sub-area of said total area, said sub-area comprising said radar plot of said aerial object,
subject said sub-area to said deep learning model on said computer system to thereby determine whether said radar plot represents an aerial object belonging to one class of aerial objects, and
conclude that said track represents an aerial object belonging to said one class when it is determined for a plurality of said radar plots in said sequence that the radar plot subjected to said deep learning model belongs to said one class.
16. The computer system according to claim 15 , wherein said deep learning model is a trained deep learning model.
17. Use of the computer system according to claim 15 to determine if an aerial object represented in a radar plot in a digital radar image is an unmanned aerial vehicle.