Object detecting apparatus, object detecting method, and computer program product

1. An object detecting apparatus, comprising:

a detector configured to detect an object candidate area from a captured image; and

an estimation processing unit configured to, by performing estimation processing using a part of or a whole of the captured image including at least the object candidate area, output object information including at least information representing a pose of an object in the object candidate area, wherein

the estimation processing includes:

a first process estimating, from among vertices of a cuboid circumscribing the object, and the vertices of the cuboid indicating positions at which the object makes contact with a road surface, positions of at least two vertices on the captured image that are viewable from a viewpoint of the captured image, to estimate anchor point positions of the object; and

a second process estimating based on positions of the vertices of the cuboid to which one of right-front, left-front, right-rear, or left-rear of the object the vertices having positions estimated by the first process are respectively corresponding, to estimate an orientation of the object, wherein

for performing the estimation processing, at least one of the first process and the second process uses at least one of a first neural network trained to estimate the anchor point positions of the object and a second neural network trained to estimate the orientation of the object.

2. The apparatus according to claim 1 , wherein

the first process estimates, from among the vertices of the cuboid circumscribing the object and making contact with the road surface, positions of vertices that are viewed at left-most and right-most on the captured image.

3. The apparatus according to claim 2 , wherein

the first process further estimates a position of a vertex on the captured image that is viewed between the vertices that are viewed at left-most and right-most.

4. The apparatus according to claim 2 , wherein

the second process estimates at which one of combinations (right-front, left-front, and left-rear), (right-front and left-front), (right-rear, right-front, and left-front), (right-rear and right-front), (left-rear, right-rear, and right-front), (left-rear and right-rear), (left-front, left-rear, and right-rear), and (left-front and left-rear) of the object is a combination of the vertices having positions estimated by the first process, sequentially from the vertex positioned at left.

5. The apparatus according to claim 4 , wherein

the second process outputs a likelihood corresponding to each of the combinations, and further outputs highest one of the likelihoods as reliability.

6. The apparatus according to claim 1 , wherein

the first process further estimates, from among the vertices of the cuboid circumscribing the object and making contact with the road surface, a position of a vertex located outside of an area of the captured image.

7. The apparatus according to claim 1 , wherein

the first process further estimates, from among the vertices of the cuboid circumscribing the object and making contact with the road surface, a position of a vertex hidden behind another object that is different from the object.

8. The apparatus according to claim 1 , wherein

the detector is configured to detect the object candidate area from the captured image, using a co-occurrence histograms of oriented gradients (CoHOG) feature.

9. The apparatus according to claim 1 , wherein

the detector is configured to detect the object candidate area using a neural network receiving an input of the captured image and outputting the object candidate area.

10. The apparatus according to claim 1 , wherein

the captured image is an image captured by an onboard camera, and

the object is at least one of a vehicle, a pedestrian, a two-wheeled vehicle, and a roadside object.

11. The apparatus according to claim 10 , wherein

the onboard camera is a monocular camera.

12. The apparatus according to claim 10 , further comprising a vehicle control unit configured to control a vehicle on which the onboard camera is mounted, based on the object information.

13. An object detecting method comprising:

detecting an object candidate from a captured image; and

outputting, by performing estimation processing using a part of or a whole of the captured image including at least the object candidate area, object information including at least information representing a pose of an object in the object candidate area, wherein

the estimation processing includes:

a first process estimating, from among vertices of a cuboid circumscribing the object, and the vertices of the cuboid indicating positions at which the object makes contact with a road surface, positions of at least two vertices on the captured image that are viewable from a viewpoint of the captured image, to estimate anchor point positions of the object; and

a second process estimating based on positions of the vertices of the cuboid to which one of right-front, left-front, right-rear, or left-rear of the object the vertices having positions estimated by the first process are respectively corresponding, to estimate an orientation of the object, wherein

for performing the estimation processing, at least one of the first process and the second process uses at least one of a first neural network trained to estimate the anchor point positions of the object and a second neural network trained to estimate the orientation type of the object.

14. A computer program product having a non-transitory computer readable medium including programmed instructions, wherein the instructions, when executed by a computer, cause the computer to perform:

a function of detecting an object candidate area from a captured image; and

a function of outputting, by performing estimation processing using a part of or a whole of the captured image including at least the object candidate area, object information including at least information representing a pose of an object in the object candidate area, wherein

the estimation processing includes:

a first process estimating, from among vertices of a cuboid circumscribing the object, and the vertices of the cuboid indicating positions at which the object makes contact with a road surface, positions of at least two vertices on the captured image that are viewable from a viewpoint of the captured image, to estimate anchor point positions of the object; and

a second process estimating based on positions of the vertices of the cuboid to which one of right-front, left-front, right-rear, or left-rear of the object the vertices having positions estimated by the first process are respectively corresponding to estimate an orientation of the object, wherein

for performing the estimation processing, at least one of the first process and the second process uses at least one of a first neural network trained to estimate the anchor point positions of the object and a second neural network trained to estimate the orientation of the object.