Detecting boxes

1. A method comprising:

receiving, at data processing hardware, a plurality of image frame pairs for an area of interest over a period of time, the area of interest comprising at least one target box, each image frame pair of the plurality of image frame pairs associated with a respective time stamp during the period of time and comprising a respective monocular image frame and a respective depth image frame; and

for each image frame pair:

determining, by the data processing hardware, two-dimensional coordinate locations for corners of a rectangle corresponding to the at least one target box within the respective monocular image frame;

based on the determined two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame:

performing, by the data processing hardware, edge detection on the at least one target box within the respective monocular image frame;

determining, by the data processing hardware, faces of the at least one target box within the respective monocular image frame; and

extracting, by the data processing hardware, planes from the respective depth image frame, the planes extracted from the respective depth image frame corresponding to the at least one target box;

matching, by the data processing hardware, the determined faces of the at least one target box within the respective monocular image frame to the planes extracted from the respective depth image frame; and

generating, by the data processing hardware, a box estimation based on the determined corners, the performed edge detection, and the matched faces.

2. The method of claim 1 , further comprising aggregating, by the data processing hardware, the box estimation generated for each image frame pair to determine an adjusted box estimation over the period of time for the at least one target box.

3. The method of claim 2 , wherein aggregating the box estimation generated for each image frame pair comprises:

comparing a first box estimation of the at least one target box generated for a first image frame pair to a second box estimation of the at least one target box generated for a second image frame pair; and

adjusting the box estimation over the period of time based on the comparison between the first box estimation and the second box estimation.

4. The method of claim 1 , further comprising displaying, by the data processing hardware, the box estimation as a homographic projection within a vision system for a robot.

5. The method of claim 1 , further comprising displaying, by the data processing hardware, the box estimation as a projection within a vision system for a robot, the projection representing the at least one target box with a color indicating a confidence interval for the box estimation.

6. The method of claim 1 , wherein determining the two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame comprises using a machine learning model to determine the two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame.

7. The method of claim 6 , wherein the machine learning model comprises a trained deep learning neural network.

8. The method of claim 1 , further comprising, for each image frame pair, cropping, by the data processing hardware, the respective monocular image frame and the respective depth image frame to isolate the area of interest corresponding to the at least one target boxes.

9. The method of claim 1 , further comprising, for each image frame pair, correcting, by the data processing hardware, angular distortion associated with the respective monocular image frame.

10. The method of claim 1 , wherein the determined two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame comprise initial seeds input to a gradient ascent optimizer, the gradient ascent optimizer configured to match faces and perform edge detection.

11. The method of claim 1 , wherein each respective depth image frame is obtained from a depth sensor, the depth sensor comprising one or more of a stereo camera, a scanning light-detection and ranging (LIDAR) sensor, a time-of-flight sensor, or a scanning laser-detection and ranging (LADAR) sensor.

12. The method of claim 1 , wherein the data processing hardware resides on a mobile robot within the area of interest.

13. The method of claim 1 , wherein the monocular image frames and the depth image frames are captured from one or more sensors mounted on an articulated arm of a robot.

14. A robot comprising:

a sensor system;

data processing hardware; and

memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:

receiving, from the sensor system, a plurality of image frame pairs for an area of interest over a period of time, the area of interest comprising at least one target box, each image frame pair of the plurality of image frame pairs associated with a respective time stamp during the period of time and comprising a respective monocular image frame and a respective depth image frame; and

for each image frame pair:

determining two-dimensional coordinate locations for corners of a rectangle corresponding to the at least one target box within the respective monocular image frame;

based on the determined two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame:

performing edge detection on the at least one target box within the respective monocular image frame;

determining faces of the at least one target box within the respective monocular image frame; and

extracting planes from the respective depth image frame, the planes extracted from the respective depth image frame corresponding to the at least one target box;

matching the determined faces of the at least one target box within the respective monocular image frame to the planes extracted from the respective depth image frame; and

generating a box estimation for the at least one target box based on the determined corners, the performed edge detection, and the matched faces.

15. The robot of claim 14 , wherein the operations further comprise aggregating the box estimation generated for each image frame pair to determine an adjusted box estimation over the period of time for the at least one target box.

16. The robot of claim 15 , wherein aggregating the box estimation generated for each image frame pair comprises:

comparing a first box estimation of the at least one target box generated for a first image frame pair to a second box estimation of the at least one target box generated for a second image frame pair; and

adjusting the box estimation over the period of time based on the comparison between the first box estimation and the second box estimation.

17. The robot of claim 14 , wherein the operations further comprise displaying the box estimation as a homographic projection within a vision system for a robot.

18. The robot of claim 14 , wherein the operations further comprise displaying the box estimation as a projection within a vision system for a robot, the projection representing the at least one target box with a color indicating a confidence interval for the box estimation.

19. The robot of claim 14 , wherein determining the two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame comprises using a machine learning model to determine the two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame.

20. The robot of claim 19 , wherein the machine learning model comprises a trained deep learning neural network.

21. The robot of claim 14 , wherein the operations further comprise, for each image frame pair, cropping the respective monocular image frame and the respective depth image frame to isolate the area of interest corresponding to the at least one target box.

22. The robot of claim 14 , wherein the operations further comprise, for each image frame pair, correcting angular distortion associated with the respective monocular image frame.

23. The robot of claim 14 , wherein the determined two-dimensional coordinate locations for corners of the rectangle within the respective monocular image frame comprise initial seeds input to a gradient ascent optimizer, the gradient ascent optimizer configured to match faces and perform edge detection.

24. The robot of claim 14 , wherein each respective depth image frame is obtained from a depth sensor of the sensor system, the depth sensor comprising one or more of a stereo camera, a scanning light-detection and ranging (LIDAR) sensor, a time-of-flight sensor, or a scanning laser-detection and ranging (LADAR) sensor.

25. The robot of claim 14 , further comprising an articulated arm, wherein at least a portion of the sensor system is mounted on the articulated arm.