Object identification apparatus, moving body system, object identification method, object identification model learning method, and object identification model learning apparatus

1. An object identification apparatus that is communicably connected to a capture portion mounted on a moving body and identifies an object in an outside of the moving body, the object identification apparatus comprising:

an image acquisition portion acquiring an image of the outside captured by the capture portion from a predetermined capture viewpoint;

a viewpoint conversion map generation portion that

forms a convolutional neural network receiving data of the image acquired by the image acquisition portion and

outputs a viewpoint conversion map obtained by converting the image into a different viewpoint from the capture viewpoint via the convolutional neural network,

wherein:

the viewpoint conversion map generation portion includes

a feature extraction process portion

applying convolution calculation by the convolutional neural network to the data of the image and

extracting a feature map of the object in a first coordinate system based on the capture viewpoint and

an output process portion

applying a warp function to the feature map extracted by the feature extraction process portion, the warp function relating a position in a second coordinate system based on the different viewpoint to a position in the first coordinate system and

outputting the viewpoint conversion map in which the object in an area of the second coordinate system is identified;

the output process portion includes

a warp application portion applying the warp function to the feature map extracted by the feature extraction process portion and

an identification process portion

concatenating the feature map to which the warp function is applied with the convolutional neural network and

outputting the viewpoint conversion map in which the object in the area of the second coordinate system is identified;

the feature extraction process portion connects a plurality of feature extraction units in series;

the plurality of feature extraction units include a convolution layer that performs the convolution calculation and a pooling layer downsampling an output result from the convolution layer by using pooling;

the identification process portion connects a plurality of identification units in series;

the plurality of identification units include an upsampling layer that performs upsampling and a deconvolution layer that applies deconvolution calculation to an output result from the upsampling layer;

the feature map to which the warp function is applied is input to the upsampling layer; and

the upsampling layer concatenates an output result from the deconvolution layer connected to an input side with the feature map to which the warp function is applied and perform output.

2. The object identification apparatus according to claim 1 , wherein:

the second coordinate system is a coordinate system of a two dimensional space having a movable direction of the moving body.

3. The object identification apparatus according to claim 1 , further comprising:

a memory device storing a kernel parameter for a kernel of the convolutional neural network and a warp parameter of the warp function; and

a calculation circuit processing the convolution calculation using the kernel parameter and calculation that uses the warp parameter and applies the warp function.

4. The object identification apparatus according to claim 1 , wherein:

the feature extraction process portion outputs the feature map in which the object in an area of the first coordinate system is identified; and

the output process portion

applies the warp function to the feature map in which the object in the area of the first coordinate system is identified and

outputs the feature map as the viewpoint conversion map.

5. The object identification apparatus according to claim 1 , wherein:

the capture portion corresponds to a camera; and

the image acquisition portion and the viewpoint conversion map generation portion correspond to a processor.

6. A moving body system for a moving body, the moving body system comprising:

a capture portion that

is mounted on the moving body and

captures an outside of the moving body from a predetermined capture viewpoint and generates an image; and

an object identification apparatus that

is communicably connected to the capture portion and

identifies an object in the outside of the moving body,

wherein:

the object identification apparatus includes

a viewpoint conversion map generation portion that

forms a convolutional neural network that receives data of the image and

outputs a viewpoint conversion map obtained by converting the image into a viewpoint different from the capture viewpoint;

the viewpoint conversion map generation portion includes

a feature extraction process portion

applying convolution calculation by the convolutional neural network to the data of the image and

extracting a feature map of the object in a first coordinate system based on the capture viewpoint, and

an output process portion

applying a warp function to the feature map extracted by the feature extraction process portion, the warp function relating a position in a second coordinate system based on the different viewpoint to a position in the first coordinate system and

outputting the viewpoint conversion map in which the object in an area of the second coordinate system is identified;

the output process portion includes

a warp application portion applying the warp function to the feature map extracted by the feature extraction process portion and

an identification process portion

concatenating the feature map to which the warp function is applied with the convolutional neural network and

outputting the viewpoint conversion map in which the object in the area of the second coordinate system is identified;

the feature extraction process portion connects a plurality of feature extraction units in series;

the plurality of feature extraction units include a convolution layer that performs the convolution calculation and a pooling layer downsampling an output result from the convolution layer by using pooling;

the identification process portion connects a plurality of identification units in series;

the plurality of identification units include an upsampling layer that performs upsampling and a deconvolution layer that applies deconvolution calculation to an output result from the upsampling layer;

the feature map to which the warp function is applied is input to the upsampling layer; and

the upsampling layer concatenates an output result from the deconvolution layer connected to an input side with the feature map to which the warp function is applied and perform output.

7. The moving body system according to claim 6 , further comprising:

an image display portion displaying an image obtained by visualizing the viewpoint conversion map.

8. The moving body system according to claim 6 , further comprising:

a movement control portion controlling movement of the moving body by using the viewpoint conversion map.

9. The moving body system according to claim 6 , wherein:

the capture portion corresponds to a camera;

the object identification, the viewpoint conversion map generation portion, the feature extraction process portion, and the output process portion correspond to a processor.

10. An object identification method comprising:

inputting data of a capture image of an object captured from a capture viewpoint to a convolutional neural network;

applying the data of the capture image to convolution calculation;

extracting a feature map in a first coordinate system based on the capture viewpoint;

applying a warp function to the feature map, the warp function relating a position in a second coordinate system based on a different viewpoint from the capture viewpoint to a position in the first coordinate system;

obtaining a viewpoint conversion map in which the data of the capture image is converted from the capture viewpoint to the different viewpoint based on the feature map to which the warp function is applied and the object is identified;

concatenating the feature map to which the warp function is applied with the convolutional neural network;

outputting the viewpoint conversion map in which the object in the area of the second coordinate system is identified;

connecting a plurality of feature extraction units in series, wherein the plurality of feature extraction units include a convolution layer that performs the convolution calculation and a pooling layer downsampling an output result from the convolution layer by using pooling;

connecting a plurality of identification units in series, wherein the plurality of identification units include an upsampling layer that performs upsampling and a deconvolution layer that applies deconvolution calculation to an output result from the upsampling layer;

inputting the feature map to which the warp function is applied to the upsampling layer; and

causing the upsampling layer to concatenate an output result from the deconvolution layer connected to an input side with the feature map to which the warp function is applied and perform output.

11. An object identification model learning method comprising:

in an object identification model forming a convolutional neural network and a warp structure warping a feature map extracted in the convolutional neural network to a different coordinate system,

preparing, in the warp structure, a warp parameter for relating a position in the different coordinate system to a position in a coordinate system before warp;

learning the warp parameter to input a capture image in which an object is captured to the object identification model and output a viewpoint conversion map in which the object is identified in the different coordinate system,

preparing depth data indication a depth of the object on the capture image; and

before the learning, determining a displacement amount of a part of the warp parameter by referring the depth data,

wherein:

the displacement amount corresponds to a position outside a blind angle area in accordance with a blind angle due to the object on the capture image; and

in the learning, in a state where the displacement amount of the part is confirmed, a kernal parameter for the convolutional neural network and a different part of the warp parameter in a non-determined state are simultaneously learned.

12. The object identification mode learning method according to claim 11 , further comprising:

before the learning, preparing a set of data of the capture image in which the object is captured and data that is correct answer data corresponding to the data of the capture image and is data of the viewpoint conversion map in which the object is identified,

wherein:

in the learning, when the data of the capture image is input to the object identification model, the kernel parameter for the convolutional neural network and the warp parameter are simultaneously learned to output data closer to the correct data is output.

13. The object identification model learning method according to claim 11 , wherein:

in the learning, a correspondence between a position on the viewpoint conversion map corresponding to the blind angle area and a position on the capture image is searched.

14. An object identification model learning apparatus for learning an object identification model, the object identification model learning apparatus comprising:

a calculation circuit executing a calculation process of the object identification model that forms a convolutional neural network and a warp structure that warps a feature map extracted in the convolutional neural network to a different coordinate system;

a teacher data setting portion setting data of a capture image of an object captured from a capture viewpoint and an output map, as correct answer data, in which the object is identified in a coordinate system based on a different viewpoint from the capture viewpoint;

a learning portion

learning a kernel parameter for a kernel of the convolutional neural network and a warp parameter for the warp structure to output data closer to the correct data when the capture image is input to the object identification model, and

before learning of the kernal parameter, determining a displacement amount of a part of the warp parameter by referring depth data indicating a depth of the object on the capture image; and

a memory device storing kernel parameter and the warp parameter that are learned by the learning portion,

wherein

the displacement amount corresponds to a position outside a blind angle area in accordance with a blind angle due to the object on the capture image; and

in the learning, in a state where the displacement amount of the part is confirmed, the kernal parameter for the convolutional neural network and a different part of the warp parameter in a non-determined state are simultaneously learned.

15. The object identification model learning apparatus according to claim 14 , wherein:

the teacher data setting portion and the learning portion correspond to a processor.

16. An object identification apparatus that is communicably connected to a camera mounted on a moving body and identifies an object in an outside of the moving body, the object identification apparatus comprising:

an image acquisition portion that is connected to the camera and acquires an image of the outside captured by the camera;

a learning value memory storing a learning value;

a calculation device

reading a kernel parameter from the learning value memory,

forming a convolutional neural network including an encoder that includes a plurality of feature amount extraction units and a decoder portion that includes a plurality of identification units,

causing the encoder portion to extract a feature map of a feature amount of the object from data of the image acquired by the image acquisition portion,

reading a warp parameter from the learning value memory,

generating a plurality of warp functions corresponding to the plurality of identification units,

applying the plurality of warp functions to the feature map to cause the plurality of identification units to read the plurality of warp functions respectively corresponding identification units, and

generating a viewpoint conversion map of which viewpoint is converted into a different viewpoint from a viewpoint captured by the camera; and

an identification process portion

concatenating the feature map to which the warp function is applied with the convolutional neural network and

outputting the viewpoint conversion map in which the object is identified, wherein

the encoder connects a plurality of feature amount extraction units in series;

the plurality of feature amount extraction units include a convolution layer that performs convolution calculation and a pooling layer downsampling an output result from the convolution layer by using pooling;

the identification process portion connects a plurality of identification units in series;

the plurality of identification units include an upsampling layer that performs upsampling and a deconvolution layer that applies deconvolution calculation to an output result from the upsampling layer;

the feature map to which the warp function is applied is input to the upsampling layer; and

the upsampling layer concatenates an output result from the deconvolution layer connected to an input side with the feature map to which the warp function is applied and perform output.

17. The object identification apparatus according to claim 16 , wherein:

the calculation device, the image acquisition portion, the encoder portion, the plurality of feature amount extraction units, the decoder portion, the plurality of identification units, and the identification process portion correspond to a processor.