System and method for estimating box shape representation of a generally cuboidal object
A system and method for estimating a 3D box model of a generally cuboidal 3D object imaged by 3D vision system, which provides a 3D image having a set of 3D points representing surfaces of the cuboidal 3D object to a processor is provided. An input process provides an approximate box as a region of interest (ROI), to the processor, the ROI defining a search volume within the 3D image. An identification process identifies the 3D points that are within the search volume. A coarse estimation process estimates 3D box dimensions that approximate a box shape of the cuboidal object based upon the identified 3D points. A refinement process refines the coarse box shape by processing the 3D points based on the 3D box dimensions that correspond to each of a plurality of imaged faces of the cuboidal object to derive an estimated result for the cuboidal object.
1 . A system for estimating a three-dimensional (3D) box model of a generally cuboidal object imaged by a 3D vision system, that provides a 3D image having 3D points representing imaged surfaces of the cuboidal object to at least one processor, comprising:
an input process that provides, to the at least one processor, a region of interest (ROI) defining a search volume within the 3D image;
an identification process that identifies the 3D points that are within the search volume;
a coarse estimation process that estimates, based on the identified 3D points, 3D box dimensions that approximate a coarse box shape of the cuboidal object; and
a refinement process that refines the coarse box shape to derive an estimated result for the 3D box model by, for an imaged surface of the imaged surfaces:
determining a projection two-dimensional (2D) height map and defining at least one of face pixels or non-face pixels associated with the imaged surface,
identifying bounding positions that define a face of the coarse box shape based upon a distribution of the at least one of face pixels or non-face pixels,
refining the face of the coarse box shape by determining an associated face thickness, and
deriving the estimated result for the 3D box model including the refined face.
2 . The system as set forth in claim 1 , wherein the coarse estimation process is adapted to estimate a thickness of each of a plurality of imaged faces corresponding to the 3D box dimensions by analyzing statistics of a distribution of the 3D points based on the 3D box dimensions that correspond to each of the plurality of imaged faces.
3 . The system as set forth in claim 2 , wherein the statistics are based upon a 3D point density distribution along a normal direction to each of the plurality of imaged faces.
4 . The system as set forth in claim 3 , wherein the statistics are further based upon a histogram of face point distances relative to each of the plurality of imaged faces, and
wherein the thickness is a predetermined distance on each of opposing sides of a histogram peak position, the histogram peak position being a location at which the 3D point density distribution is a maximum.
5 . The system as set forth in claim 2 , wherein the refinement process is adapted to identify boundary 3D points corresponding to the 3D points located between adjacent faces of the plurality of imaged faces.
6 . The system as set forth in claim 5 , wherein the refinement process is adapted to compute a face plane for each of the plurality of imaged faces by fitting a 3D robust plane using the identified boundary 3D points.
7 . The system as set forth in claim 6 , wherein the face plane is computed based upon the thickness estimates of neighboring imaged faces.
8 . The system as set forth in claim 1 , further comprising a result process that is adapted to estimate a 3D box shape based upon the 3D box dimensions that correspond to a plurality of imaged faces with face position and rotation correction.
9 . The system as set forth in claim 8 , wherein the result process is adapted to perform rotation correction based upon identification of a face in the estimated result that defines a tilt angle that is greater than a preset threshold angle.
10 . A method for estimating a three-dimensional (3D) box model of a generally cuboidal object imaged by a 3D vision system that provides a 3D image having 3D points representing imaged surfaces of the cuboidal object, the method comprising the steps of:
providing a region of interest (ROI) in which the ROI defines a search volume within the 3D image;
identifying the 3D points that are within the search volume;
estimating 3D box dimensions that approximate a coarse box shape of the cuboidal object based upon the identified 3D points;
refining the coarse box shape by determining a projection two-dimensional (2D) height map associated with an imaged surface of the imaged surfaces, and defining at least one of face pixels or non-face pixels associated with the imaged surface;
identifying bounding positions that define a face of the coarse box shape based upon a distribution of the at least one of face pixels or non-face pixels;
refining the face of the coarse box shape by determining an associated face thickness; and
deriving an estimated result for the 3D box model including the refined face.
11 . The method as set forth in claim 10 , wherein the step of estimating determines an estimated thickness of each of a plurality of imaged faces corresponding to the 3D box dimensions by analyzing statistics of a distribution of the 3D points based on the 3D box dimensions that correspond to each of the plurality of imaged faces.
12 . The method as set forth in claim 11 , further comprising, basing the statistics upon a 3D point density distribution along a normal direction to each of the plurality of imaged faces.
13 . The method as set forth in claim 12 , further comprising, basing the statistics upon a histogram of face point distances relative to each of the plurality of imaged faces, and
determining the thickness based upon a predetermined distance on each of opposing sides of a histogram peak position, the histogram peak position being a location at which the 3D point density distribution is a maximum.
14 . The method as set forth in claim 11 , wherein the step of refining identifies boundary 3D points corresponding to the 3D points located between adjacent faces of the plurality of imaged faces.
15 . The method as set forth in claim 14 , wherein the step of refining comprises computing a face plane for each of the plurality of imaged faces by fitting a 3D robust plane using the identified boundary 3D points.
16 . The method as set forth in claim 15 , wherein the step of computing the face plane is based upon the thickness estimate of neighboring imaged faces.
17 . The method as set forth in claim 10 , further comprising, estimating a 3D box shape based upon the 3D box dimensions that correspond to a plurality of imaged faces with face position and rotation correction.
18 . The method as set forth in claim 17 , further comprising, correcting rotation of the estimated result based upon identification of a face in the estimated result that defines a tilt angle that is greater than a preset threshold angle.
19 . A non-transitory computer readable medium comprising program instructions that, when executed, cause at least one processor to perform:
receiving, by the at least one processor, a three-dimensional (3D) image of a generally cuboidal object, the 3D image having 3D points representing imaged surfaces of the cuboidal object;
receiving, by the at least one processor, a region of interest (ROI) in which the ROI defines a search volume within the 3D image;
identifying the 3D points that are within the search volume;
estimating 3D box dimensions that approximate a coarse box shape of the cuboidal object based upon the identified 3D points;
refining the coarse box shape by determining a projection two-dimensional (2D) height map associated with an imaged surface of the imaged surfaces, and defining at least one of face pixels or non-face pixels associated with the imaged surface;
identifying bounding positions that define a face of the coarse box shape based upon a distribution of the at least one of face pixels or non-face pixels;
refining the face of the coarse box shape by determining an associated face thickness; and
deriving an estimated result for the 3D box model including the refined face.
20 . The non-transitory computer readable medium as set forth in claim 19 , further comprising program instructions that, when executed, cause the at least one processor to determine an estimated thickness of each of a plurality of imaged faces corresponding to the 3D box dimensions by analyzing statistics of a distribution of the 3D points based on the 3D box dimensions that correspond to each of the plurality of imaged faces.