Image depth prediction method, electronic device, and non-transitory storage medium

1 . An image depth prediction method, comprising:

acquiring image frames of containing a dynamic object by a monocular camera and extracting a continuous of object frames and reference frames from the image frames;

reconstructing the object frames to obtain reconstructed frames according to the reference frames and a preset depth estimation model, and obtaining a reconstruction error between the object frames and the reconstructed frames;

processing the image frames to obtain point cloud data and instance segmentation data;

fusing the point cloud data with the instance segmentation data to obtain mask data;

obtaining a loss function according to the reconstruction error and the mask data and training the preset depth estimation model based on the loss function until the loss function converges to optimize the preset depth estimation model; and

spreading forward the preset depth estimation model to calculate a predictive depth image of new object frames.

2 . The image depth prediction method of claim 1 , wherein processing the image frames to obtain the point cloud data and the instance segmentation data comprises:

obtaining instance segmentation data of each pixel in the image frames based on an instance segmentation network; and

obtaining point cloud data of a laser lidar by using the laser lidar to scan the image frames.

3 . The image depth prediction method of claim 2 , wherein fusing the point cloud data with the instance segmentation data to obtain the mask data comprises:

acquiring an extrinsic parameter of the monocular camera and the laser lidar, wherein the extrinsic parameter comprises a rotation matrix and a translation matrix;

projecting the point cloud data of the laser lidar into a three-dimensional coordinate system of the monocular camera according to the external parameter;

acquiring an internal parameter of the monocular camera, wherein the internal parameter comprises an internal parameter matrix and a distortion parameter matrix;

projecting points of the monocular camera in the three-dimensional coordinate system onto an imaging plane according to the internal parameter and obtaining a mapping relationship between the point cloud data of the laser lidar and the pixels of the image frames; and

attaching the instance segmentation data of each pixel in the image frames to the point cloud data according to the correspondence mapping relationship to obtain the mask data.

4 . The image depth prediction method of claim 3 , wherein the calculation formula in the mapping relationship between the point cloud data of the laser lidar and the pixels of the image frames comprises:

ZP

uv

=

Z

[

u

v

1

]

=

[

f

x

0

c

x

0

f

y

c

y

0

0

1

]

[

X

Y

Z

]

=

KP

;

where P is a three-dimensional coordinate of the point cloud data, Z is a z component of the three-dimensional coordinates of the point cloud data, P uv is a two-dimensional coordinate of the imaging plane, X is a x component of the three-dimensional coordinates of the point cloud data, Y is a y component of the three-dimensional coordinates of the point cloud data, K is an internal parameter of the monocular camera.

5 . The image depth prediction method of claim 1 , wherein reconstructing the object frames to obtain the reconstructed frames according to the reference frames and the preset depth estimation model comprises:

performing a depth estimation on the object frames through the preset depth estimation model to obtain depth information of the object frames;

inputting the object frames and the reference frames into a preset pose estimation model to obtain a camera pose change between the object frames and the reference frames; and

reconstructing the object frames to obtain reconstructed frames corresponding to the object frames by the depth information and the camera pose change.

6 . The image depth prediction method of claim 1 , wherein obtaining the reconstruction error between the object frames and the reconstructed frames comprises:

calculating a luminance difference of the object frames and the reconstruction frames to obtain the reconstruction error.

7 . An electronic device, comprising:

at least one processor; and

a data storage storing one or more programs which when executed by the at least one processor, cause the at least one processor to:

acquire image frames of containing a dynamic object by a monocular camera and extract a continuous of object frames and reference frames from the image frames;

reconstruct the object frames to obtain reconstructed frames according to the reference frames and a preset depth estimation model, and obtain a reconstruction error between the object frames and the reconstructed frames;

process the image frames to obtain point cloud data and instance segmentation data;

fuse the point cloud data with the instance segmentation data to obtain mask data;

obtain a loss function according to the reconstruction error and the mask data and train the preset depth estimation model based on the loss function until the loss function converges; and

spread forward the preset depth estimation model to calculate a predictive depth image of a new object frame.

8 . The electronic device of claim 7 , wherein the at least one processor processing the image frames to obtain the point cloud data and the instance segmentation data comprises:

obtaining instance segmentation data of each pixel in the image frames based on an instance segmentation network; and

obtaining point cloud data of a laser lidar by using the laser lidar to scan the image frames.

9 . The electronic device of claim 8 , wherein the at least one processor fusing the point cloud data with the instance segmentation data to obtain the mask data comprises:

acquiring an extrinsic parameter of the monocular camera and the laser lidar, wherein the extrinsic parameter comprises a rotation matrix and a translation matrix;

projecting the point cloud data of the laser lidar into a three-dimensional coordinate system of the monocular camera according to the external parameter;

acquiring an internal parameter of the monocular camera, wherein the internal parameter comprises an internal parameter matrix and a distortion parameter matrix;

projecting points of the monocular camera in the three-dimensional coordinate system onto an imaging plane according to the internal parameter and obtaining a mapping relationship between the point cloud data of the laser lidar and the pixels of the image frames; and

attaching the instance segmentation data of each pixel in the image frames to the point cloud data according to the correspondence mapping relationship to obtain the mask data.

10 . The electronic device of claim 9 , wherein the calculation formula in the mapping relationship between the point cloud data of the laser lidar and the pixels of the image frames comprises:

ZP

uv

=

Z

[

u

v

1

]

=

[

f

x

0

c

x

0

f

y

c

y

0

0

1

]

[

X

Y

Z

]

=

KP

;

where P is a three-dimensional coordinate of the point cloud data, Z is a z component of the three-dimensional coordinates of the point cloud data, P uv is a two-dimensional coordinate of the imaging plane, X is a x component of the three-dimensional coordinates of the point cloud data, Y is a y component of the three-dimensional coordinates of the point cloud data, K is an internal parameter of the monocular camera.

11 . The electronic device of claim 7 , wherein the at least one processor reconstructing the object frames to obtain the reconstructed frames according to the reference frames and the preset depth estimation model comprises:

performing a depth estimation on the object frames through the preset depth estimation model to obtain depth information of the object frames;

inputting the object frames and the reference frames into a preset pose estimation model to obtain a camera pose change between the object frames and the reference frames; and

reconstructing the object frames to obtain reconstructed frames corresponding to the object frames by the depth information and the camera pose change.

12 . The electronic device of claim 7 , wherein obtaining the reconstruction error between the object frames and the reconstructed frames comprises:

calculating a luminance difference of the object frames and the reconstruction frames to obtain the reconstruction error.

13 . A non-transitory storage medium having stored thereon instructions that, when executed by a processor of an electronic device, causes the electronic device to perform an image depth prediction method, the image depth prediction method comprises:

acquiring image frames of containing a dynamic object by a monocular camera and extracting a continuous of object frames and reference frames from the image frames;

reconstructing the object frames to obtain reconstructed frames according to the reference frames and a preset depth estimation model, and obtaining a reconstruction error between the object frames and the reconstructed frames;

processing the image frames to obtain point cloud data and instance segmentation data;

fusing the point cloud data with the instance segmentation data to obtain mask data;

obtaining a loss function according to the reconstruction error and the mask data and training the preset depth estimation model based on the loss function until the loss function converges; and

spreading forward the preset depth estimation model to calculate a predictive depth image of a new object frame.

14 . The non-transitory storage medium of claim 13 , wherein processing the image frames to obtain the point cloud data and the instance segmentation data comprises:

obtaining instance segmentation data of each pixel in the image frames based on an instance segmentation network; and

obtaining point cloud data of a laser lidar by using the laser lidar to scan the image frames.

15 . The non-transitory storage medium of claim 14 , wherein fusing the point cloud data with the instance segmentation data to obtain the mask data comprises:

acquiring an extrinsic parameter of the monocular camera and the laser lidar, wherein the extrinsic parameter comprises a rotation matrix and a translation matrix;

projecting the point cloud data of the laser lidar into a three-dimensional coordinate system of the monocular camera according to the external parameter;

acquiring an internal parameter of the monocular camera, wherein the internal parameter comprises an internal parameter matrix and a distortion parameter matrix;

projecting points of the monocular camera in the three-dimensional coordinate system onto an imaging plane according to the internal parameter and obtaining a mapping relationship between the point cloud data of the laser lidar and the pixels of the image frames; and

attaching the instance segmentation data of each pixel in the image frames to the point cloud data according to the correspondence mapping relationship to obtain the mask data.

16 . The non-transitory storage medium of claim 15 , wherein the calculation formula in the mapping relationship between the point cloud data of the laser lidar and the pixels of the image frames comprises:

ZP

uv

=

Z

[

u

v

1

]

=

[

f

x

0

c

x

0

f

y

c

y

0

0

1

]

[

X

Y

Z

]

=

KP

;

where P is a three-dimensional coordinate of the point cloud data, Z is a z component of the three-dimensional coordinates of the point cloud data, P uv is a two-dimensional coordinate of the imaging plane, X is a x component of the three-dimensional coordinates of the point cloud data, Y is a y component of the three-dimensional coordinates of the point cloud data, K is an internal parameter of the monocular camera.

17 . The non-transitory storage medium of claim 13 , wherein reconstructing the object frames to obtain the reconstructed frames according to the reference frames and the preset depth estimation model comprises:

performing a depth estimation on the object frames through the preset depth estimation model to obtain depth information of the object frames;

inputting the object frames and the reference frames into a preset pose estimation model to obtain a camera pose change between the object frames and the reference frames; and

reconstructing the object frames to obtain reconstructed frames corresponding to the object frames by the depth information and the camera pose change.

18 . The non-transitory storage medium of claim 13 , wherein obtaining the reconstruction error between the object frames and the reconstructed frames comprises:

calculating a luminance difference of the object frames and the reconstruction frames to obtain the reconstruction error.