Method and system for scanning an object using an RGB-D sensor

1. A method for scanning an object comprising:

generating a plurality of elementary scans of the object using an RGB-D sensor, each elementary scan including a plurality of frames;

using each of the plurality of frames of the respective elementary scans to obtain poses of the scan via visual odometry, and for each elementary scan selecting a plurality of key frames from the plurality of frames with respect to the object, wherein each elementary scan performed with the object is in a different position relative to the sensor;

for each elementary scan, rotating the object or rotating the sensor around the object, the object being in the position associated with the respective elementary scan;

for each elementary scan, elaborating a three-dimensional model of the object using the plurality of key frames and poses of the scan; and

merging each three-dimensional model into a merged three-dimensional model of the object, wherein

the merging comprises determining transformations that transform said each three-dimensional model into an initial three-dimensional model,

the determining the transformations comprises aligning said each three-dimensional model into the initial three-dimensional model using an approximate L1 norm, and

the elaborating the three-dimensional model for each elementary scan is made using a variational technique using colored signed distance fields and an approximate L 1 minimization.

2. The method according to claim 1 , wherein loop closure is detected after a revolution of the object or a revolution of the sensor around the object.

3. The method according to claim 2 , wherein a pose graph optimization is carried out on the poses of each elementary scan after detecting loop closure.

4. The method according to claim 1 , further comprising using the marching cube algorithm on the merged three-dimensional model of the object to obtain a three-dimensional mesh of the object.

5. The method according to claim 1 , wherein positions of the object during at least two elementary scans display different portions of the object to the sensor.

6. The method according to claim 1 , further comprising a segmentation of each key frame to separate the object from the background on the key frame.

7. A system for scanning an object, comprising:

an RGB-D sensor configured to be directed towards the object;

means for rotating the object or rotating the sensor around the object; and

means for receiving and processing data from the RGB-D sensor for a plurality of elementary scans of the object, each elementary scan including a plurality of frames;

the means for receiving and processing data

using each of the plurality of frames of the respective elementary scans to obtain poses of the scan via visual odometry,

selecting a plurality of key frames from the plurality of frames with respect to the object, wherein each elementary scan performed with the object is in a different position relative to the sensor,

elaborating a three-dimensional model of the object using the plurality of key frames and poses of the scan of the object obtained by the rotating, and

merging a plurality of three-dimensional models into a merged three-dimensional model of the object, each of the plurality of three-dimensional models scanned using the RGB-D sensor at a respective different position of the object, wherein

the merging comprises determining transformations that transform each of the plurality of three-dimensional models into an initial three-dimensional model,

the determining the transformations comprises aligning said each three-dimensional model into the initial three-dimensional model using an approximate L1 norm, and

the elaborating the three-dimensional model for each elementary scan is made using a variational technique using colored signed distance fields and an approximate L 1 minimization.

8. The system according to claim 7 , further comprising means for detecting loop closure after a revolution of the object or a revolution of the sensor around the object.

9. The system according to claim 8 , further comprising means for carrying out a pose-graph optimization on the poses of each elementary scan after detecting loop closure.

10. A system for scanning an object, comprising:

an RGB-D sensor configured to be directed towards the object;

a rotating table or robot arm that rotates the object; and

processing circuitry configured to

receive and process data from the RGB-D sensor for a plurality of elementary scans of the object, each elementary scan including a plurality of frames,

use each of the plurality of frames of the respective elementary scans to obtain poses of the scan via visual odometry,

select a plurality of key frames from the plurality of frames with respect to the object, wherein each elementary scan performed with the object is in a different position relative to the sensor,

elaborate a three-dimensional model of the object using the plurality of key frames and poses of the scan of the object obtained by the rotating, and

merge a plurality of three-dimensional models into a merged three-dimensional model of the object, each of the plurality of three-dimensional models scanned using the RGB-D sensor at a respective different position of the object, wherein

the merge comprises determining transformations that transform each of the plurality of three-dimensional models into an initial three-dimensional model,

the determining the transformations comprises aligning said each three-dimensional model into the initial three-dimensional model using an approximate L1 norm, and

the elaborate the three-dimensional model for each elementary scan is made using a variational technique using colored signed distance fields and an approximate L 1 minimization.