End-to-end multi-person articulated three dimensional pose tracking

1. An end-to-end method of articulated three-dimensional pose tracking comprising:

receiving, from one or more cameras, a plurality of frames of image data;

extracting first feature data representing the plurality of frames of image data from a first backbone convolutional neural network (CNN);

projecting the first feature data into three-dimensional (3D) feature data in a voxel space;

detecting a first person represented in one or more of the plurality of frames of image data using a 3D CNN;

generating location data describing a location of the first person in the voxel space;

sending the first feature data and the location data to a four-dimensional (4D) CNN effective to generate second feature data comprising a 3D representation of the first person over time;

generating merged feature data based at least in part on merging the second feature data over time; and

performing one or more deconvolution operations on the merged feature data to generate 3D pose information for the first person over time.

2. The method of claim 1 , further comprising:

determining, for the first person represented in the second feature data, spatial contextual information using an attention mechanism; and

determining, for the first person represented in the second feature data, temporal contextual information using the attention mechanism, wherein the merged feature data comprises the spatial contextual information and the temporal contextual information.

3. The method of claim 1 , further comprising:

determining, using the second feature data, a similarity between a first feature representation of the first person at time t and a second feature representation of the first person at time t+Δt;

determining, using the similarity, that the first feature representation and the second feature representation correspond to the same person; and

generating matching data indicating that the first feature representation and the second feature representation correspond to the same person, wherein the merged feature data comprises combined representations of feature representations of the same person.

4. A method comprising:

receiving a plurality of frames of image data captured by one or more cameras;

determining first feature data representing the plurality of frames of image data using a backbone network;

projecting the first feature data into three-dimensional (3D) space;

determining 3D location data describing respective 3D locations of one or more persons represented by the first feature data projected in the 3D space;

sending the first feature data and the 3D location data to a four-dimensional (4D) convolutional neural network (CNN);

generating, by the 4D CNN, second feature data comprising respective 3D representations of the one or more persons based on context from frames of the plurality of frames representing different points in time; and

generating 3D pose data representing articulated 3D pose information for the one or more persons.

5. The method of claim 4 , further comprising:

determining, using the second feature data, a similarity between a first feature representation of a person at time t and a second feature representation of the person at time t+Δt;

determining, using the similarity, that the first feature representation and the second feature representation correspond to the same person; and

generating matching data indicating that the first feature representation and the second feature representation correspond to the same person, wherein the second feature data comprises combined representations of feature representations of the same person.

6. The method of claim 4 , further comprising:

determining the 3D location data using a 3D CNN; and

determining a person detection loss representing differences between ground truth 3D locations of one or more persons represented in the plurality of frames of image data and the 3D location data.

7. The method of claim 4 , further comprising generating merged feature data based at least in part on merging the second feature data over multiple time steps represented by the plurality of frames of image data.

8. The method of claim 7 , further comprising generating the 3D pose data representing articulated 3D pose information for the one or more persons over time based at least in part on performing one or more deconvolution operations on the merged feature data.

9. The method of claim 8 , further comprising:

determining a pose prediction loss representing differences between the 3D pose data and ground truth 3D pose data; and

updating parameters of the backbone network and the 4D CNN based at least in part on the pose prediction loss.

10. The method of claim 4 , further comprising training at least the backbone network, the 4D CNN, and a 3D person detection CNN configured to generate the 3D location data end-to-end.

11. The method of claim 10 , further comprising:

training at least the backbone network, the 4D CNN, and the 3D person detection CNN end-to-end using a person tracking loss related to tracking of persons in the 3D space over time and a pose estimation loss related to the articulated 3D pose information for the one or more persons over time.

12. The method of claim 4 , further comprising:

determining, for a first person represented in the second feature data, spatial contextual information using an attention mechanism; and

determining, for the first person represented in the second feature data, temporal contextual information using the attention mechanism, wherein the 3D pose data representing articulated 3D pose information for the one or more persons is generated based at least in part on the spatial contextual information and the temporal contextual information.

13. A system comprising:

at least one processor; and

at least one non-transitory computer-readable memory storing instructions that, when executed by the at least one processor, are effective to program the at least one processor to:

receive a plurality of frames of image data captured by one or more cameras;

determine first feature data representing the plurality of frames of image data using a backbone network;

project the first feature data into three-dimensional (3D) space;

determine 3D location data describing respective 3D locations of one or more persons represented by the first feature data projected in the 3D space;

send the first feature data and the 3D location data to a four-dimensional (4D) convolutional neural network (CNN);

generate, by the 4D CNN, second feature data comprising respective 3D representations of the one or more persons; and

generate 3D pose data representing articulated 3D pose information for the one or more persons.

14. The system of claim 13 , the at least one non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to program the at least one processor to:

determine, using the second feature data, a similarity between a first feature representation of a person at time t and a second feature representation of the person at time t+Δt;

determine, using the similarity, that the first feature representation and the second feature representation correspond to the same person; and

generate matching data indicating that the first feature representation and the second feature representation correspond to the same person, wherein the second feature data comprises combined representations of feature representations of the same person.

15. The system of claim 13 , the at least one non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to program the at least one processor to:

determine the 3D location data using a 3D CNN; and

determine a person detection loss representing differences between ground truth 3D locations of one or more persons represented in the plurality of frames of image data and the 3D location data.

16. The system of claim 13 , the at least one non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to program the at least one processor to generate merged feature data based at least in part on merging the second feature data over multiple time steps represented by the plurality of frames of image data.

17. The system of claim 16 , the at least one non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to program the at least one processor to generate the 3D pose data representing articulated 3D pose information for the one or more persons over time based at least in part on performing one or more deconvolution operations on the merged feature data.

18. The system of claim 17 , the at least one non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to program the at least one processor to:

determine a pose prediction loss representing differences between the 3D pose data and ground truth 3D pose data; and

update parameters of the backbone network and the 4D CNN based at least in part on the pose prediction loss.

19. The system of claim 13 , the at least one non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to program the at least one processor to train at least the backbone network, the 4D CNN, and a 3D person detection CNN configured to generate the 3D location data end-to-end.

20. The system of claim 19 , the at least one non-transitory computer-readable memory storing further instructions that, when executed by the at least one processor, are further effective to train at least the backbone network, the 4D CNN, and the 3D person detection CNN end-to-end using a person tracking loss related to tracking of persons in the 3D space over time and a pose estimation loss related to the articulated 3D pose information for the one or more persons over time.