Processing method of a 3D point cloud

1. A method of processing a three-dimensional point cloud, comprising:

obtaining a 3D point cloud from a predetermined view point of a depth sensor;

extracting 3D coordinates and intensity data from each point of the 3D point cloud with respect to the view point;

transforming 3D coordinates and intensity data into at least three two-dimensional (2D) spaces, utilizing:

an intensity 2D space function of the intensity data of each point;

a height 2D space function of an elevation data of each point, and

a distance 2D space function of a distance data between each point of 3D point cloud and the view point; and

defining a single multi-channel 2D space based on the at least three two-dimensional (2D) spaces.

2. The method of processing method of claim 1 , wherein the transforming step further includes:

detecting background points; and

setting a predetermined default value for detected background points.

3. The method of processing method of claim 1 , further including a training phase that further includes before the transforming step:

supplying tags data classifying objects in 1:N classes; and

labelling each extracted point data as belonging to objects of 1:N classes according to the supplied tags data.

4. The method of processing method of claim 1 , wherein the transforming step includes a normalizing step of at least one of the 2D spaces, which includes:

normalizing the intensity space by applying an intensity normalizing function depending on the distance data to the intensity data of each point; and/or

normalizing the height space by applying a height normalizing function depending on the minimum and maximum elevation data within the 3D point cloud to the elevation data of each point; and/or

normalizing the distance space by applying a distance normalizing function depending on the minimum and maximum distance data within the 3D point cloud to the distance data of each point, and

defining a normalized single multi-channel 2D space.

5. The method of processing method of claim 1 , wherein the 3D point cloud is derived using the depth sensor such as a Lidar.

6. The method of processing method of claim 1 , wherein each of the intensity space, the height space and the distance space are respectively assigned to one of RGB value spaces for each point.

7. The processing method of claim 2 , further including a training phase that further includes before the transforming step:

supplying tags data classifying objects in 1:N classes; and

labelling each extracted point data as belonging to objects of 1:N classes according to the supplied tags data.

8. The processing method of claim 2 , wherein the transforming step includes a normalizing step of at least one of the 2D spaces, which includes:

normalizing the intensity space by applying an intensity normalizing function depending on the distance data to the intensity data of each point; and/or

normalizing the height space by applying a height normalizing function depending on the minimum and maximum elevation data within the 3D point cloud to the elevation data of each point; and/or

normalizing the distance space by applying a distance normalizing function depending on the minimum and maximum distance data within the 3D point cloud to the distance data of each point, and

defining a normalized single multi-channel 2D space.

9. The processing method of claim 2 , wherein the 3D point cloud is derived using the depth sensor such as a Lidar.

10. The processing method of claim 2 , wherein each of the intensity space, the height space and the distance space are respectively assigned to one of RGB value spaces for each point.

11. The processing method of claim 3 , wherein the transforming step includes a normalizing step of at least one of the 2D spaces, which includes:

normalizing the intensity space by applying an intensity normalizing function depending on the distance data to the intensity data of each point; and/or

normalizing the height space by applying a height normalizing function depending on the minimum and maximum elevation data within the 3D point cloud to the elevation data of each point; and/or

normalizing the distance space by applying a distance normalizing function depending on the minimum and maximum distance data within the 3D point cloud to the distance data of each point, and

defining a normalized single multi-channel 2D space.

12. The processing method of claim 3 , wherein the 3D point cloud is derived using the depth sensor such as a Lidar.

13. The processing method of claim 3 , wherein each of the intensity space, the height space and the distance space are respectively assigned to one of RGB value spaces for each point.

14. The processing method of claim 4 , wherein the 3D point cloud is derived using the depth sensor such as a Lidar.

15. The processing method of claim 4 , wherein each of the intensity space, the height space and the distance space are respectively assigned to one of RGB value spaces for each point.

16. A method for classification of an object of interest within a 3D point cloud, comprising:

the processing method according to claim 1 ;

supplying at least one single multi-channel 2D space based on the at least three two-dimensional (2D) spaces to a machine learning such as a deep neural network; and

outputting at least one classification label for each single multi-channel 2D space based on the at least three two-dimensional (2D) spaces supplied identifying at least one object of interest within the 3D point cloud.

17. The classification method of claim 16 , wherein the machine learning comprises a training phase during which the supplying step includes:

accumulating a batch of multi-channel 2D spaces by repeating the processing method according to claim 1 ; and

supplying an average combination of a random number of multi-channel 2D spaces among said batch of accumulated multi-channel 2D spaces.

18. The classification method of claim 16 , wherein the machine learning comprises a training phase during which the supplying step includes:

accumulating a batch of multi-channel 2D spaces by repeating the processing method according to claim 1 ; and

supplying several multi-channel 2D spaces distributed in a predefined order among said batch of accumulated multi-channel 2D spaces.

19. The classification method of claim 16 , wherein the machine learning comprises a training phase during which the supplying step includes:

accumulating a batch of multi-channel 2D spaces by repeating the processing method according to claim 1 ; and

supplying a matrix of at least 2×2 multi-channel 2D spaces randomly selected among said batch of accumulated multi-channel 2D spaces.

20. A vehicle, comprising:

a depth sensor configured to obtain 3D point clouds of a followed path when the vehicle is driven at least from a driver's point of view,

a processor to process the obtained 3D point clouds that is configured to operate according to the processing method of claim 1 ;

a classifying unit to classify an object of interest within processed 3D point clouds;

a display configured to display an information related to the classified object of interest;

an autonomous driving unit configured to control the vehicle depending on the classified object of interest; and

a decision unit to decide to activate at least one of the display unit and the autonomous driving unit.