Machine-learned monocular depth estimation and semantic segmentation for 6-DOF absolute localization of a delivery drone
A method includes receiving a two-dimensional (2D) image captured by a camera on a unmanned aerial vehicle (UAV) and representative of an environment of the UAV. The method further includes applying a trained machine learning model to the 2D image to produce a semantic image of the environment and a depth image of the environment, where the semantic image comprises one or more semantic labels. The method additionally includes retrieving reference depth data representative of the environment, wherein the reference depth data includes reference semantic labels. The method also includes aligning the depth image of the environment with the reference depth data representative of the environment to determine a location of the UAV in the environment, where the aligning associates the one or more semantic labels from the semantic image with the reference semantic labels from the reference depth data.
1. A method comprising:
receiving a two-dimensional (2D) image captured by a camera on a unmanned aerial vehicle (UAV) and representative of an environment of the UAV;
applying a trained machine learning model to the 2D image to produce a semantic image of the environment and a depth image of the environment, wherein the machine learning model has been trained with a semantics branch to produce the semantic image and a depth branch to produce the depth image, and wherein the semantic image comprises one or more semantic labels;
retrieving reference depth data representative of the environment, wherein the reference depth data includes reference semantic labels; and
determining a location of the UAV in the environment, wherein determining the location of the UAV in the environment comprises aligning the depth image of the environment with the reference depth data representative of the environment, wherein aligning the depth image of the environment with the reference depth data representative of the environment is based on associating the one or more semantic labels from the semantic image with the reference semantic labels from the reference depth data; and
controlling the UAV to navigate in the environment based on the determined location of the UAV in the environment.
2. The method of claim 1 , further comprising:
controlling the UAV to navigate in the environment using a Global Navigation Satellite System (GNSS) system;
detecting a disruption in service from the GNSS system, wherein the location of the UAV in the environment is determined responsive to detecting the disruption in service from the GNSS system; and
subsequent to detecting the disruption in service from the GNSS system, controlling the UAV to navigate in the environment based on the determined location of the UAV in the environment.
3. The method of claim 1 , further comprising:
controlling the UAV to navigate in the environment using a GNSS system; and
using the determined location of the UAV in the environment to cross-check location data from the GNSS system.
4. The method of claim 1 , further comprising:
determining a GNSS location of the UAV in the environment using a GNSS system;
determining a refined location of the UAV in the environment based on the GNSS location of the UAV in the environment and the determined location of the UAV in the environment; and
controlling the UAV to navigate in the environment based on the refined location of the UAV in the environment.
5. The method of claim 1 , wherein the aligning the depth image of the environment with the reference depth data representative of the environment to determine the location of the UAV in the environment comprises using an iterative closest point (ICP) algorithm.
6. The method of claim 5 , wherein the ICP algorithm aligns points from the reference depth data with points from the depth image such that the reference semantic labels from the reference depth data correspond to the one or more semantic labels from the semantic image.
7. The method of claim 1 , wherein the semantic image of the environment and the depth image of the environment produced by the machine learning model have the same dimensions.
8. The method of claim 1 , wherein the camera on the UAV faces downward, and wherein the 2D image captured by the camera is representative of a terrain in the environment below the UAV.
9. The method of claim 1 , wherein the semantics branch and the depth branch of the machine learning model operate on a commonly generated feature set.
10. The method of claim 1 , wherein the machine learning model has been trained based on ground truth depth data, wherein the ground truth depth data is based on performance of a structure from motion (SfM) algorithm on images captured by one or more UAVs.
11. The method of claim 1 , wherein the machine learning model has been trained based on ground truth semantic data, wherein the ground truth semantic data is based on operator labeling of images captured by one or more UAVs.
12. The method of claim 1 , wherein the machine learning model has been trained using a scale invariant loss for training of the depth branch.
13. The method of claim 12 , further comprising applying a scale factor to the depth image, wherein the scale factor comprises a ratio of an altitude of the UAV above ground level relative to a median of a monocular depth map, wherein the monocular depth map is based on the reference depth data.
14. The method of claim 12 , further comprising applying a scale factor to the depth image, wherein the scale factor comprises a ratio of an altitude of the UAV above ground level relative to an above ground level estimate from a monocular depth map, wherein the monocular depth map is based on the reference depth data.
15. The method of claim 1 , wherein the one or more semantic labels are selected from a predetermined set of labels, wherein the predetermined set of labels comprises at least the following labels: building, road, vegetation, vehicle, driveway, lawn, and sidewalk.
16. The method of claim 1 , further comprising retrieving the reference depth data in advance of a flight of the UAV, wherein the reference depth data is selected based on a planned flight path of the UAV.
17. The method of claim 1 , further comprising applying a Kalman filter to the determined location of the UAV in the environment to control navigation of the UAV in the environment.
18. An unmanned aerial vehicle (UAV), comprising:
a camera; and
a control system configured to:
receive a two-dimensional (2D) image captured by the camera and representative of an environment of the UAV;
apply a trained machine learning model to the 2D image to produce a semantic image of the environment and a depth image of the environment, wherein the machine learning model has been trained with a semantics branch to produce the semantic image and a depth branch to produce the depth image, and wherein the semantic image comprises one or more semantic labels;
retrieve reference depth data representative of the environment, wherein the reference depth data includes reference semantic labels; and
determine a location of the UAV in the environment, wherein determining the location of the UAV in the environment comprises aligning the depth image of the environment with the reference depth data representative of the environment, wherein aligning the depth image of the environment with the reference depth data representative of the environment is based on associating the one or more semantic labels from the semantic image with the reference semantic labels from the reference depth data.
19. A non-transitory computer readable medium comprising program instructions executable by one or more processors to perform operations, the operations comprising:
receiving a two-dimensional (2D) image captured by a camera on a unmanned aerial vehicle (UAV) and representative of an environment of the UAV;
applying a trained machine learning model to the 2D image to produce a semantic image of the environment and a depth image of the environment, wherein the machine learning model has been trained with a semantics branch to produce the semantic image and a depth branch to produce the depth image, and wherein the semantic image comprises one or more semantic labels;
retrieving reference depth data representative of the environment, wherein the reference depth data includes reference semantic labels; and
determining a location of the UAV in the environment, wherein determining the location of the UAV in the environment comprises aligning the depth image of the environment with the reference depth data representative of the environment, wherein aligning the depth image of the environment with the reference depth data representative of the environment is based on associating the one or more semantic labels from the semantic image with the reference semantic labels from the reference depth data; and
controlling the UAV to navigate in the environment based on the determined location of the UAV in the environment.