Image-based depth data and relative depth data
A vehicle can use an image sensor to both detect objects and determine depth data associated with the environment the vehicle is traversing. The vehicle can capture image data and lidar data using the various sensors. The image data can be provided to a machine-learned model trained to output depth data of an environment. Such models may be trained, for example, by using lidar data and/or three-dimensional map data associated with a region in which training images and/or lidar data were captured as ground truth data. The autonomous vehicle can further process the depth data and generate additional data including localization data, three-dimensional bounding boxes, and relative depth data and use the depth data and/or the additional data to autonomously traverse the environment, provide calibration/validation for vehicle sensors, and the like.
1. A system comprising:
one or more processors; and
one or more computer-readable media storing computer-executable instructions that, when executed, cause the one or more processors to perform operations comprising:
capturing, by an image sensor of an autonomous vehicle, image data associated with an environment;
determining an object represented in the image data;
capturing, by a lidar sensor of the autonomous vehicle, lidar data associated with the object, a portion of the lidar data associated with the object comprising a reference point;
inputting a crop of the image data and the reference point to a machine-learned model;
receiving, from the machine-learned model, relative depth data associated with the object, the relative depth data comprising depth data that is relative to a depth associated with the reference point; and
determining, based at least in part on the reference point and the relative depth data, estimated depth data.
2. The system of claim 1 , the operations further comprising:
segmenting the image data to identify a portion of the image data associated with the object; and
determining, based at least in part on segmenting the image data, the crop of the image data.
3. The system of claim 2 , the operations further comprising:
determining, based at least in part on the image data, depth data; and
determining that the crop of the image data is associated with sparse depth data.
4. The system of claim 1 , the operations further comprising:
determining, based at least in part on the image data, depth data;
determining a confidence level associated with a portion of the depth data; and
determining, based at least in part on the confidence level, the reference point.
5. The system of claim 1 , wherein the machine-learned model is trained based at least in part on a plurality of images, an image of the plurality of images comprising a training object associated with a plurality of lidar measurements.
6. The system of claim 1 , the operations further comprising:
determining, based at least in part on the estimated depth data, one or more of dimensions associated with the object or an estimated pose of the object.
7. A method comprising:
receiving, from an image sensor on a vehicle, image data;
determining an object represented in the image data;
receiving depth data representing a portion of an environment, the depth data comprising sparse depth data;
associating a portion of the depth data with the object, wherein the portion of the depth data and comprises a reference point;
inputting a portion of the image data associated with the object and the reference point to a machine-learned model;
receiving, from the machine-learned model, relative depth data associated with the object, the relative depth data comprising a plurality of depths relative to a depth associated with the reference point; and
determining, based at least in part on the depth associated with the reference point and the relative depth data, estimated depth data associated with the object.
8. The method of claim 7 ,
wherein the machine-learned model is a first machine-learned model, and
wherein determining the portion of the image data associated with the object comprises:
inputting the image data into a second machine-learned model trained to detect objects; and
receiving, from the second machine-learned model, the portion of the image data.
9. The method of claim 7 ,
wherein the portion of the depth data is a second portion of second depth data captured at a second time,
the method further comprising:
receiving first depth data at a first time, the first time preceding the second time;
determining a first portion of the first depth data associated with the object, a first number of measurements of the first portion of the first depth data being greater than a second number of measurements of the second portion of the second depth data;
determining, based at least in part on the first portion of the first depth data, first object parameters;
determining, based at least in part on the estimated depth data, second object parameters; and
associating, as an object track, first object parameters and second object parameters.
10. The method of claim 8 , the method further comprising:
determining, based at least in part on the image data, depth data;
determining a confidence level associated with a second portion of the depth data; and
determining, based at least in part on the confidence level, the reference point.
11. The method of claim 7 , wherein the machine-learned model is trained based at least in part on captured image data and captured depth data corresponding to at least a second portion of the captured image data, and wherein the captured depth data represents ground truth data for training the machine-learned model, and wherein training the machine-learned model comprises substantially minimizing a loss based on a difference between a depth value output by the machine-learned model and a ground truth depth value based on the captured depth data.
12. The method of claim 7 , wherein the sparse depth data is received from a lidar sensor.
13. The method of claim 7 , further comprising:
determining, based at least in part on the reference point, an image coordinate associated with the reference point,
wherein the estimated depth data is further based at least in part on the image coordinate.
14. A non-transitory computer-readable medium storing instructions executable by a processor, wherein the instructions, when executed, cause the processor to perform operations comprising:
receiving, from an image sensor on a vehicle, image data;
determining an object represented in the image data;
receiving depth data representing a portion of an environment, the depth data comprising sparse depth data;
associating a portion of the depth data with the object, wherein the portion of the depth data and comprises a reference point;
inputting a portion of the image data associated with the object and the reference point to a machine-learned model;
receiving, from the machine-learned model, relative depth data associated with the object, the relative depth data comprising a plurality of depths relative to a depth associated with the reference point; and
determining, based at least in part on the depth associated with the reference point and the relative depth data, estimated depth data associated with the object.
15. The non-transitory computer-readable medium of claim 14 ,
wherein the machine-learned model is a first machine-learned model, and
wherein determining the portion of the image data associated with the object comprises:
inputting the image data into a second machine-learned model trained to detect objects; and
receiving, from the second machine-learned model, the portion of the image data.
16. The non-transitory computer-readable medium of claim 15 ,
wherein the portion of the depth data is a second portion of second depth data captured at a second time,
the operations further comprising:
receiving first depth data at a first time, the first time differing from the second time;
determining a first portion of the first depth data associated with the object, a first number of measurements of the first portion of the first depth data being greater than a second number of measurements of the second portion of the second depth data;
determining, based at least in part on the first portion of the first depth data, first object parameters;
determining, based at least in part on the estimated depth data, second object parameters; and
associating, as an object track, first object parameters and second object parameters.
17. The non-transitory computer-readable medium of claim 15 , the operations further comprising:
determining, based at least in part on the image data, depth data;
determining a confidence level of the depth data associated with the object; and
determining, based at least in part on the confidence level being less than or equal to a threshold confidence level, the reference point.
18. The non-transitory computer-readable medium of claim 14 , wherein the machine-learned model is trained based at least in part on captured image data and captured depth data corresponding to at least a second portion of the captured image data, and wherein the captured depth data represents ground truth data for training the machine-learned model, and wherein training the machine-learned model comprises substantially minimizing a loss based on a difference between a depth value output by the machine-learned model and a ground truth depth value based on the captured depth data.
19. The non-transitory computer-readable medium of claim 14 , wherein the sparse depth data is received from a lidar sensor.
20. The non-transitory computer-readable medium of claim 14 , the operations further comprising:
determining, based at least in part on the reference point, an image coordinate associated with the reference point,
wherein the estimated depth data is further based at least in part on the image coordinate.