Apparatus and method for object detection
View Patent ↗An object detection apparatus includes a sensor device and a processor. The sensor device obtains an image. The processor obtains a first bounding box and a second bounding box based on a learning result of the image, calculates a two-dimensional overlapping degree between the first bounding box and the second bounding box, calculates a depth deviation between the first bounding box and the second bounding box, calculates a three-dimensional overlapping degree by reflecting the depth deviation to the two-dimensional overlapping degree, and performs a non-maximum suppression procedure based on the three-dimensional overlapping degree.
1 . An object detection apparatus comprising:
a sensor device configured to obtain an image; and
a processor configured to
obtain a first bounding box and a second bounding box based on a learning result of the image,
calculate a two-dimensional overlapping degree of between the first bounding box and the second bounding box,
calculate a depth deviation between the first bounding box and the second bounding box, and
perform a non-maximum suppression procedure based on a three-dimensional overlapping degree,
wherein the processor is further configured to:
obtain a normalization factor by calculating a difference between “1” and a normalized depth deviation; and
obtain the three-dimensional overlapping degree by performing a multiplication operation between the two-dimensional overlapping degree and the normalization factor.
2 . The object detection apparatus of claim 1 , wherein the processor is configured to select the first bounding box and the second bounding box within output values of a network using an anchor box corresponding to two or more vehicle sizes.
3 . The object detection apparatus of claim 2 , wherein the processor is configured to:
calculate the two-dimensional overlapping degree of between the first bounding box and the second bounding box;
determine a two-dimensional region of the first bounding box and a two-dimensional region of the second bounding box on an x-y plane of an input image; and
calculate an overlapping degree between the two-dimensional region of the first bounding box and the two-dimensional region of the second bounding box, and
wherein an x-axis of the x-y plane is perpendicular to a traveling direction of a vehicle on a road surface, and wherein a y-axis of the x-y plane is an axis perpendicular to the road surface.
4 . The object detection apparatus of claim 3 , wherein the processor is configured to obtain the depth deviation by calculating a difference in coordinate values on a z-axis between a reference point of the first bounding box and a reference point of the second bounding box on a bird's eye view of the image including the z-axis perpendicular to the x-y plane.
5 . The object detection apparatus of claim 4 , wherein the processor is configured to calculate the three-dimensional overlapping degree based on an absolute value of the depth deviation.
6 . The object detection apparatus of claim 5 , wherein the processor is configured to normalize the depth deviation within a range greater than 0 and less than 1, based on a ratio of the depth deviation to a maximum deviation of a depth value.
7 . The object detection apparatus of claim 6 , wherein the processor is configured to:
calculate an average depth value for each type of the anchor box; and
obtain a maximum deviation of the depth value by calculating a difference between a maximum value of the average depth value and a minimum value of the average depth value.
8 . The object detection apparatus of claim 6 , wherein the processor is configured to set a size of the three-dimensional overlapping degree to be small in inverse proportion to the depth deviation.
9 . The object detection apparatus of claim 1 , wherein the processor is configured to exclude a bounding box having a low confidence among the first bounding box and the second bounding box from an object candidate group when the three-dimensional overlapping degree is equal to or greater than a preset threshold value.
10 . An object detection method comprising:
obtaining a first bounding box and a second bounding box based on a learning result of an input image;
calculating a two-dimensional overlapping degree between the first bounding box and the second bounding box;
calculating a depth deviation between the first bounding box and the second bounding box; and
performing a non-maximum suppression procedure based on a three-dimensional overlapping degree,
wherein calculating the three-dimensional overlapping degree includes:
obtaining a normalization factor by calculating a difference between “1” and a normalized depth deviation; and
obtaining the three-dimensional overlapping degree by performing a multiplication operation between the two-dimensional overlapping degree and the normalization factor.
11 . The object detection method of claim 10 , wherein obtaining the first bounding box and the second bounding box includes selecting a pair of bounding boxes within output values of a network using an anchor box corresponding to two or more vehicle sizes.
12 . The object detection method of claim 11 , wherein
the input image includes a plane defined by an x-axis and a y-axis;
calculating the two-dimensional overlapping degree between the first bounding box and the second bounding box includes
determining a two-dimensional region of the first bounding box and a two-dimensional region of the second bounding box on an x-y plane, and
calculating an overlapping degree between the two-dimensional region of the first bounding box and the two-dimensional region of the second bounding box;
the x-axis of the x-y plane is perpendicular to a traveling direction of a vehicle on a road surface; and
the y-axis of the x-y plane is an axis perpendicular to the road surface.
13 . The object detection method of claim 12 , wherein
calculating the depth deviation includes obtaining the depth deviation by calculating a difference in coordinate values on a z-axis between a reference point of the first bounding box and a reference point of the second bounding box on a bird's eye view of the input image, and
the z-axis is perpendicular to the x-y plane.
14 . The object detection method of claim 13 , wherein obtaining the depth deviation further includes obtaining an absolute value of the depth deviation.
15 . The object detection method of claim 14 , wherein obtaining the depth deviation further includes normalizing the depth deviation within a range greater than 0 and less than 1, based on a ratio of the depth deviation to a maximum deviation of a depth value.
16 . The object detection method of claim 15 , wherein obtaining the depth deviation further includes:
calculating an average depth value for each of types of the anchor box; and
obtaining a maximum deviation of the depth value by calculating a difference between a maximum value of the average depth value and a minimum value of the average depth value.
17 . The object detection method of claim 15 , wherein calculating the three-dimensional overlapping degree includes setting a size of the three-dimensional overlapping degree to be small in inverse proportion to the depth deviation.
18 . The object detection method of claim 10 , wherein performing the non-maximum suppression procedure includes excluding a bounding box having a low confidence among the first bounding box and the second bounding box from an object candidate group when the three-dimensional overlapping degree is equal to or greater than a preset threshold value.