Estimating three-dimensional target heading using a single snapshot

1. A computer implemented method performed by one or more processors, comprising:

obtaining a three-dimensional (3D) snapshot of radar data comprising Doppler velocities and spatial positions of a plurality of detection points of a target, wherein the obtaining of the 3D snapshot comprises:

in response to a numerical quantity of detection points along a dimension being less than a threshold number of points, fusing, using a convolutional neural network (CNN), remaining detection points along other two dimensions with Lidar data along the dimension to obtain the Doppler velocities and the spatial positions of the remaining detection points;

conducting a first estimation of a 3D heading of the target based on the spatial positions;

conducting a second estimation of the 3D heading of the target based on the Doppler velocities; and

obtaining a combined estimation of the 3D heading of the target based on a weighted sum of the first estimation and the second estimation.

2. The computer implemented method of claim 1 , further comprising determining a first weight associated with the first estimation of the weighted sum and a second weight associated with the second estimation of the weighted sum based on respective first and second weights, obtained from a previous cycle, at a location within a threshold distance of at least a portion of the spatial positions.

3. The computer implemented method of claim 1 , wherein the conducting the second estimation is further based on a least-squares solution of a velocity vector of the target.

4. The computer implemented method of claim 1 , wherein the conducting the second estimation is further based on directions from each of the detection points to a radar sensor used to obtain the 3D snapshot and magnitudes of Doppler speeds at each of the detection points.

5. The computer implemented method of claim 1 , further comprising:

determining, based on the first estimation, a 3D bounding region enclosing the detection points, the 3D bounding region indicating an orientation and a dimension of the target.

6. The computer implemented method of claim 5 , wherein the conducting the first estimation of the 3D heading comprises determining a cuboid bounding region that minimizes a sum of distances from each detection point to a surface of the cuboid bounding region.

7. The computer implemented method of claim 6 , wherein the distances are determined from each detection point to a nearest surface of the determined cuboid bounding region.

8. The computer implemented method of claim 5 , further comprising determining a first weight associated with the first estimation of the weighted sum and a second weight associated with the second estimation of the weighted sum based on a variance of the combined estimation.

9. The computer implemented method of claim 5 , wherein the snapshot comprising the radar data is obtained from multiple radar sensors; and

the determining the 3D boundary region is based on the radar data obtained from the multiple radar sensors.

10. The computer implemented method of claim 5 , further comprising:

obtaining a second snapshot of Lidar data comprising a plurality of second detection points within second threshold distances of at least a portion of the detection points;

determining, based on the second snapshot, a second 3D bounding region enclosing the second detection points;

determining whether boundaries of the second 3D bounding region are within third threshold distances of surfaces of the 3D bounding region; and

in response to determining that the surfaces of the second 3D bounding region are within third threshold distances of the surfaces of the 3D bounding region, fusing the 3D bounding region and the second 3D bounding region.

11. The computer implemented method of claim 5 , further comprising:

obtaining a third snapshot of camera data comprising a plurality of third detection points within third threshold distances of at least a portion of the detection points;

determining, based on the third snapshot, a third 3D bounding region enclosing the third detection points;

determining whether boundaries of the third 3D bounding region are within fourth threshold distances of surfaces of the 3D bounding region; and

in response to determining that the surfaces of the third 3D bounding region are within fourth threshold distances of the surfaces of the 3D bounding region, fusing the 3D bounding region and the third 3D bounding region.

12. The computer implemented method of claim 5 , wherein each of the detection points contacts a surface of the 3D bounding region or is located in an interior of the 3D bounding region.

13. A system configured to determine a three-dimensional (3D) heading of a target, comprising:

a radar sensor configured to obtain at least a portion of a 3D snapshot of radar data comprising Doppler velocities and spatial positions of a plurality of detection points of a target;

a Lidar sensor configured to capture Lidar data;

one or more processors; and

a memory storing instructions that, when executed by the one or more processors, causes the system to perform:

in response to a numerical quantity of detection points along a dimension being less than a threshold number of points, fusing, using a convolutional neural network (CNN), remaining detection points along other two dimensions with Lidar data along the dimension to obtain the Doppler velocities and the spatial positions of the remaining detection points;

conducting a first estimation of the 3D heading of the target based on the spatial positions;

conducting a second estimation of the 3D heading of the target based on the Doppler velocities; and

obtaining a combined estimation of the 3D heading of the target based on a weighted sum of the first estimation and the second estimation.

14. The system of claim 13 , wherein the instructions further cause the system to perform:

determining a first weight associated with the first estimation of the weighted sum and a second weight associated with the second estimation of the weighted sum based on respective first and second weights, obtained from a previous cycle, at a location within a threshold distance of at least a portion of the spatial positions.

15. The system of claim 13 , wherein the conducting the second estimation is further based on a least-squares solution of a velocity vector of the target.

16. The system of claim 13 , wherein the instructions further cause the system to perform:

determining, based on the first estimation, a 3D bounding region enclosing the detection points, the 3D bounding region indicating an orientation and a dimension of the target.

17. The system of claim 16 , further comprising a second radar sensor configured to obtain a second portion of the 3D snapshot comprising the radar data; and

the determining the 3D boundary region is based on the portion and the second portion of the snapshot.

18. The system of claim 16 , wherein the conducting the first estimation of the 3D heading comprises determining a cuboid bounding region that minimizes a sum of distances from each detection point to a surface of the cuboid bounding region.

19. The system of claim 18 , wherein each of the detection points contacts a surface of the cuboid bounding region or is located in an interior of the cuboid bounding region.