Method and system for automatically processing point cloud based on reinforcement learning

1. A method for automatically processing point cloud based on reinforcement learning, comprising:

scanning to collect a point cloud (PCL) and an image through a Light Detection and Ranging (Lidar) and a camera;

calibrating, by a controller, to match locations of the image and the point cloud through reinforcement learning that maximizes a reward including geometric and luminous intensity consistency of the image and the point cloud; and

meshing, by the controller, the point cloud into a 3D image through reinforcement learning that minimizes a reward including a difference between a shape of the image and a shape of the point cloud,

wherein the meshing is performed by an action including a motion vector of a catch particle, and

wherein the meshing comprises determining the motion vector to capture a point cloud having highest correlation among adjacent point clouds.

2. The method of claim 1 , wherein the calibrating is performed by an action comprising a translation velocity vector and a rotation velocity vector.

3. The method of claim 2 , wherein in the calibrating, the reward comprises a luminous intensity loss which is a pixel-wise error between a mis-calibrated depth map and a target depth map, and a point cloud distance loss which is a distance between a transformed point cloud and a target point cloud.

4. The method of claim 3 , wherein the point cloud distance loss is a sum of squares of distances of closest points between two point clouds.

5. The method of claim 2 , wherein in the calibrating, the reinforcement learning is performed by a discrete value iteration (DVI) reinforcement learning algorithm.

6. The method of claim 2 , wherein the calibrating comprises transforming the point cloud into a depth map by projecting the point cloud onto an image plane.

7. The method of claim 6 , wherein the calibrating comprises a step of 3D spatial transformation which transforms the depth map by a transformation matrix.

8. The method of claim 1 , wherein in the meshing, the reward comprises a difference between a location of an octree data extracted from the image and a location of a point cloud to be connected next which is captured by the catch particle through a motion according to the action.

9. The method of claim 1 , wherein the meshing comprises dividing the point cloud into small unit screens and parsing while moving a window.

10. The method of claim 1 , wherein in the meshing, the reinforcement learning is performed by a discrete value iteration (DVI) reinforcement learning algorithm.

11. A reinforcement learning based Light Detection and Ranging (Lidar)-camera calibration method that matches locations of a point cloud (PCL) and an image collected by a Lidar and a camera, based on reinforcement learning, in a system for automatically processing point cloud based on reinforcement learning,

wherein the reinforcement learning is performed by a reward for maximizing geometric and luminous intensity consistency of the image and the point cloud, and an action including a translation velocity vector and a rotation velocity vector, and

wherein the reward comprises a luminous intensity loss which is a pixel-wise error between a mis-calibrated depth map and a target depth map, and a point cloud distance loss which is a distance between a transformed point cloud and a target point cloud.

12. The method of claim 11 , wherein the point cloud distance loss is a sum of squares of distances of closest points between two point clouds.

13. The method of claim 11 , wherein the reinforcement learning is performed by a discrete value iteration (DVI) reinforcement learning algorithm.

14. The method of claim 11 , further comprising: transforming the point cloud into a depth map by projecting the point cloud onto an image plane.

15. The method of claim 14 , further comprising a step of 3D spatial transformation which transforms the depth map by a transformation matrix.

16. A reinforcement learning based point cloud mesh method that meshes a point cloud into a 3D image, using a point cloud and an image collected by a Light Detection and Ranging (Lidar) and a camera, based on reinforcement learning, in a system for automatically processing point cloud based on reinforcement learning,

wherein the reinforcement learning is performed by a reward for minimizing a difference between a shape of the image and a shape of the point cloud, and an action including a motion vector of a catch particle, and

wherein the reinforcement learning comprises determining the motion vector to capture a point cloud having highest correlation among adjacent point clouds.

17. The method of claim 16 , wherein the reward comprises a difference between a location of an octree data extracted from the image and a location of a point cloud to be connected next which is captured by the catch particle through a motion according to the action.

18. The method of claim 16 , further comprising:

dividing the point cloud into small unit screens and parsing while moving a window.

19. The method of claim 16 , wherein the reinforcement learning is performed by a discrete value iteration (DVI) reinforcement learning algorithm.

20. A system for automatically processing point cloud based on reinforcement learning, comprising:

a Light Detection and Ranging (Lidar) for collecting a point cloud;

a camera for collecting an image; and

a controller communicatively connected to the Lidar and the camera and configured to:

calibrate to match locations of the image and the point cloud through reinforcement learning that maximizes a reward including geometric and luminous intensity consistency of the image and the point cloud, and

mesh the point cloud into a 3D image through reinforcement learning that minimizes a reward including a difference between a shape of the image and a shape of the point cloud,

wherein the controller is configured to mesh the point cloud into the 3D image by an action including a motion vector of a catch particle, and

wherein the controller is configured to determine the motion vector to capture a point cloud having highest correlation among adjacent point clouds.

21. A reinforcement learning based Light Detection and Ranging (Lidar)-camera calibration system, comprising:

a Lidar for collecting a point cloud;

a camera for collecting an image; and

a controller communicatively connected to the Lidar and the camera and configured to match locations of the point cloud and the image based on reinforcement learning,

wherein the reinforcement learning is performed by a reward for maximizing geometric and luminous intensity consistency of the image and the point cloud, and an action including a translation velocity vector and a rotation velocity vector, and

wherein the reward comprises a luminous intensity loss which is a pixel-wise error between a mis-calibrated depth map and a target depth map, and a point cloud distance loss which is a distance between a transformed point cloud and a target point cloud.

22. A reinforcement learning based point cloud mesh processing system, comprising:

a Light Detection and Ranging (Lidar) for collecting a point cloud;

a camera for collecting an image; and

a controller communicatively connected to the Lidar and the camera and configured to mesh the point cloud into a 3D image, using the point cloud and the image, based on reinforcement learning,

wherein the reinforcement learning is performed by a reward for minimizing a difference between a shape of the image and a shape of the point cloud, and an action including a motion vector of a catch particle, and

wherein the reinforcement learning comprises determining the motion vector to capture a point cloud having highest correlation among adjacent point clouds.