Dynamic scene three-dimensional reconstruction method, apparatus and system, server, and medium

1. A three-dimensional reconstruction method of a dynamic scene, comprising:

acquiring a plurality of first continuous depth image sequences of the dynamic scene, wherein the plurality of first continuous depth image sequences are captured by an array of drones, wherein each drone is equipped with a depth camera;

fusing the plurality of first continuous depth image sequences to establish a three-dimensional reconstruction model of the dynamic scene;

determining target observation points of the array of drones according to the three-dimensional reconstruction model and current poses of the array of drones; and

instructing the array of drones to move to the target observation points to capture a plurality of second continuous depth image sequences of the dynamic scene, and updating the three-dimensional reconstruction model according to the plurality of second continuous depth image sequences captured by the array of drones at the target observation points.

2. The method of claim 1 , wherein acquiring the plurality of first continuous depth image sequences of the dynamic scene comprises:

acquiring a plurality of original depth image sequences of the dynamic scene captured by the array of drones; and

aligning the plurality of original depth image sequences according to a synchronization timestamp to obtain the plurality of first continuous depth image sequences.

3. The method of claim 1 , wherein fusing the plurality of first continuous depth image sequences to establish the three-dimensional reconstruction model of the dynamic scene comprises:

fusing the plurality of first continuous depth image sequences, determining a key frame reconstruction body according to a preset period, and performing, in each preset period, following operations:

determining a deformation parameter of a non-rigid deformation node in a current key frame reconstruction body, and updating a reconstruction model in the current key frame reconstruction body to a current data frame reconstruction body according to the deformation parameter, wherein the current data frame reconstruction body refers to a real-time reconstruction body at each moment;

extracting the three-dimensional reconstruction model of the dynamic scene from the current data frame reconstruction body; and

using the current data frame reconstruction body to replace the current key frame reconstruction body to act as a key frame reconstruction body in a next preset period.

4. The method of claim 1 , wherein determining the target observation points of the array of drones according to the three-dimensional reconstruction model and the current poses of the array of drones comprises:

rasterizing a spatial neighborhood of the array of drones according to the current poses of the array of drones to establish a set of candidate observation points;

determining a total energy value of each candidate observation point in the set of candidate observation points by using a validity energy function; and

using candidate observation points each of which the total energy value meets a preset standard as the target observation points.

5. The method of claim 4 , wherein the validity energy function comprises a depth energy term, a center energy term and a motion energy term;

wherein the depth energy term is used for determining a degree of closeness between an average depth value of each candidate observation point and a respective target depth value;

the center energy term is used for determining a degree of closeness between a reconstruction model observed at the candidate observation points and a center portion of a captured image; and

the motion energy term is used for determining an amount of moving portions in the dynamic scene observed at the candidate observation points.

6. The method of claim 5 , wherein the validity energy function is expressed by a following formula:

E t =λ d E d +λ c E c +λ m E m

wherein E t denotes a total energy term, Ed denotes the depth energy term, Ec denotes the center energy term, Em denotes the motion energy term, and λ d denotes a weight coefficient corresponding to the depth energy term, λ c denotes a weight coefficient corresponding to the center energy term, and λ m denotes a weight coefficient corresponding to the motion energy term; and

the depth energy term, the center energy term and the motion energy term are respectively expressed by following formulas:

E

d

=

ψ

⁡

(

d

a

⁢

v

⁢

g

-

d

o

)

E

c

=

∑

r

1

λ

+

d

⁢

u

2

+

1

λ

+

d

⁢

v

2

E

m

=

∑

r

ϕ

1

(

r

)

+

∑

x

i

ϕ

2

(

x

i

)

ϕ

1

(

r

)

=



1

+

N

x

′

(

T

c

-

1

⁢

t

v

-

x

n

)



2

2

ϕ

2

(

x

i

)

=



π

⁡

(

T

V

⁢

T

C

(

x

i

′

-

x

i

)

)



2

2

wherein T c denotes poses of the array of drones in the reconstruction model and T V denotes poses of the candidate observation points in the reconstruction model; t v denotes translation components of the poses of the candidate observation points; x n denotes a voxel of the reconstruction model hit by a ray; N x denotes a normal direction of the voxel; x i denotes a node to be subjected to non-rigid deformation in the reconstruction model; x i ′ denotes a node having been subjected to the non-rigid deformation; π( ) denotes a projection perspective transformation from three-dimensional space to a two-dimensional image plane; d avg denotes an average depth value of each candidate observation point and d o denotes a respective target depth value; ψ( ) denotes a penalization term for a distance; r denotes light rays projected from the candidate observation points and passing through the reconstruction model; du and dv denote an average abscissa and an average ordinate of projection pixels of the reconstruction model at each candidate observation point, respectively; λ denotes a damping factor; and ϕ1 is used for counting motion information of all rays of the candidate observation points and ϕ2 is used for counting motion information of all observed deformation nodes.

7. A three-dimensional reconstruction apparatus of a dynamic scene, comprising:

at least one processor; and

a memory communicatively connected to the at least one processor;

wherein the memory stores instructions executable by the at least one processor, and the at least one processor is configured to execute the instructions to:

acquire a plurality of first continuous depth image sequences of the dynamic scene, wherein the plurality of first continuous depth image sequences are captured by an array of drones, wherein each drone is equipped with a depth camera;

fuse the plurality of first continuous depth image sequences to establish a three-dimensional reconstruction model of the dynamic scene;

determine target observation points of the array of drones according to the three-dimensional reconstruction model and current poses of the array of drones; and

instruct the array of drones to move to the target observation points to capture a plurality of second continuous depth image sequences of the dynamic scene, and update the three-dimensional reconstruction model according to the plurality of second continuous depth image sequences captured by the array of drones at the target observation points.

8. The apparatus of claim 7 , wherein the at least one processor is configured to execute the instructions to acquire the plurality of first continuous depth image sequences of the dynamic scene in the following manners:

acquiring a plurality of original depth image sequences of the dynamic scene captured by the array of drones; and

aligning the plurality of original depth image sequences according to a synchronization timestamp to obtain the plurality of first continuous depth image sequences.

9. The apparatus of claim 7 , wherein the at least one processor is configured to execute the instructions to fuse the plurality of first continuous depth image sequences to establish the three-dimensional reconstruction model of the dynamic scene in the following manners:

fusing the plurality of first continuous depth image sequences, determining a key frame reconstruction body according to a preset period, and performing, in each preset period, following operations:

determining a deformation parameter of a non-rigid deformation node in a current key frame reconstruction body, and updating a reconstruction model in the current key frame reconstruction body to a current data frame reconstruction body according to the deformation parameter, wherein the current data frame reconstruction body refers to a real-time reconstruction body at each moment;

extracting the three-dimensional reconstruction model of the dynamic scene from the current data frame reconstruction body; and

using the current data frame reconstruction body to replace the current key frame reconstruction body to act as a key frame reconstruction body in a next preset period.

10. The apparatus of claim 7 , wherein the at least one processor is configured to execute the instructions to determine the target observation points of the array of drones according to the three-dimensional reconstruction model and the current poses of the array of drones in the following manners:

rasterizing a spatial neighborhood of the array of drones according to the current poses of the array of drones to establish a set of candidate observation points;

determining a total energy value of each candidate observation point in the set of candidate observation points by using a validity energy function; and

using candidate observation points each of which the total energy value meets a preset standard as the target observation points.

11. The apparatus of claim 10 , wherein the validity energy function comprises a depth energy term, a center energy term and a motion energy term;

wherein the depth energy term is used for determining a degree of closeness between an average depth value of each candidate observation point and a respective target depth value;

the center energy term is used for determining a degree of closeness between a reconstruction model observed at the candidate observation points and a center portion of a captured image;

and the motion energy term is used for determining an amount of moving portions in the dynamic scene observed at the candidate observation points.

12. The apparatus of claim 11 , wherein the validity energy function is expressed by a following formula:

E t =λ d E d +λ c E c +λ m E m

wherein E t denotes a total energy term, Ed denotes the depth energy term, Ec denotes the center energy term, Em denotes the motion energy term, and λ d denotes a weight coefficient corresponding to the depth energy term, λ c denotes a weight coefficient corresponding to the center energy term, and λ m denotes a weight coefficient corresponding to the motion energy term; and

the depth energy term, the center energy term and the motion energy term are respectively expressed by following formulas:

E

d

=

ψ

⁡

(

d

a

⁢

v

⁢

g

-

d

o

)

E

c

=

∑

r

1

λ

+

d

⁢

u

2

+

1

λ

+

d

⁢

v

2

E

m

=

∑

r

ϕ

1

(

r

)

+

∑

x

i

ϕ

2

(

x

i

)

ϕ

1

(

r

)

=



1

+

N

x

′

(

T

c

-

1

⁢

t

v

-

x

n

)



2

2

ϕ

2

(

x

i

)

=



π

⁡

(

T

V

⁢

T

C

(

x

i

′

-

x

i

)

)



2

2

wherein T c denotes poses of the array of drones in the reconstruction model and T V denotes poses of the candidate observation points in the reconstruction model; t v denotes translation components of the poses of the candidate observation points; x n denotes a voxel of the reconstruction model hit by a ray; N x denotes a normal direction of the voxel; x i denotes a node to be subjected to non-rigid deformation in the reconstruction model; x i ′ denotes a node having been subjected to the non-rigid deformation; π( ) denotes a projection perspective transformation from three-dimensional space to a two-dimensional image plane; d avg denotes an average depth value of each candidate observation point and d o denotes a respective target depth value; ψ( ) denotes a penalization term for a distance; r denotes light rays projected from the candidate observation points and passing through the reconstruction model; du and dv denote an average abscissa and an average ordinate of projection pixels of the reconstruction model at each candidate observation point, respectively; λ denotes a damping factor; and ϕ1 is used for counting motion information of all rays of the candidate observation points and ϕ2 is used for counting motion information of all observed deformation nodes.

13. A three-dimensional reconstruction system of a dynamic scene, comprising an array of drones and a three-dimensional reconstruction platform;

wherein each drone in the array of drones is equipped with a depth camera, and the depth camera is configured to capture a depth image sequence of the dynamic scene; and

the three-dimensional reconstruction platform comprises the three-dimensional reconstruction apparatus of the dynamic scene according to claim 7 , and is configured to generate a three-dimensional reconstruction model of the dynamic scene according to a plurality of continuous depth image sequences captured by the array of drones.

14. The system of claim 13 , wherein

the three-dimensional reconstruction platform further comprises a wireless communication module wirelessly connected to the array of drones, wherein the wireless communication module is configured to receive the plurality of continuous depth image sequences captured by the array of drones and is further configured to send position information of target observation points determined by the three-dimensional reconstruction apparatus to the array of drones; and

each drone in the array of drones further comprises a navigation module, wherein the navigation module is configured to control, according to the position information, the drone to move to a target observation point to capture the dynamic scene.

15. A server, comprising:

at least one processor; and

a storage device, which is configured to store at least one program;

wherein the at least one program, when executed by the at least one processor, causes the at least one processor to implement the three-dimensional reconstruction method of the dynamic scene of claim 1 .

16. A non-transitory computer-readable storage medium, storing a computer program, wherein the computer program, when executed by a processor, implements the three-dimensional reconstruction method of the dynamic scene of claim 1 .

17. The method of claim 2 , wherein fusing the plurality of first continuous depth image sequences to establish the three-dimensional reconstruction model of the dynamic scene comprises:

fusing the plurality of first continuous depth image sequences, determining a key frame reconstruction body according to a preset period, and performing, in each preset period, following operations:

determining a deformation parameter of a non-rigid deformation node in a current key frame reconstruction body, and updating a reconstruction model in the current key frame reconstruction body to a current data frame reconstruction body according to the deformation parameter, wherein the current data frame reconstruction body refers to a real-time reconstruction body at each moment;

extracting the three-dimensional reconstruction model of the dynamic scene from the current data frame reconstruction body; and

using the current data frame reconstruction body to replace the current key frame reconstruction body to act as a key frame reconstruction body in a next preset period.

18. The apparatus of claim 8 , wherein the at least one processor is configured to execute the instructions to fuse the plurality of first continuous depth image sequences to establish the three-dimensional reconstruction model of the dynamic scene in the following manners:

fusing the plurality of first continuous depth image sequences, determining a key frame reconstruction body according to a preset period, and performing, in each preset period, following operations:

determining a deformation parameter of a non-rigid deformation node in a current key frame reconstruction body, and updating a reconstruction model in the current key frame reconstruction body to a current data frame reconstruction body according to the deformation parameter, wherein the current data frame reconstruction body refers to a real-time reconstruction body at each moment;

extracting the three-dimensional reconstruction model of the dynamic scene from the current data frame reconstruction body; and

using the current data frame reconstruction body to replace the current key frame reconstruction body to act as a key frame reconstruction body in a next preset period.

19. A three-dimensional reconstruction system of a dynamic scene, comprising an array of drones and a three-dimensional reconstruction platform;

wherein each drone in the array of drones is equipped with a depth camera, and the depth camera is configured to capture a depth image sequence of the dynamic scene; and

the three-dimensional reconstruction platform comprises the three-dimensional reconstruction apparatus of the dynamic scene according to claim 8 , and is configured to generate a three-dimensional reconstruction model of the dynamic scene according to a plurality of continuous depth image sequences captured by the array of drones.

20. A three-dimensional reconstruction system of a dynamic scene, comprising an array of drones and a three-dimensional reconstruction platform;

wherein each drone in the array of drones is equipped with a depth camera, and the depth camera is configured to capture a depth image sequence of the dynamic scene; and

the three-dimensional reconstruction platform comprises the three-dimensional reconstruction apparatus of the dynamic scene according to claim 9 , and is configured to generate a three-dimensional reconstruction model of the dynamic scene according to a plurality of continuous depth image sequences captured by the array of drones.