Machine learning techniques for ground classification
Example systems, methods, and non-transitory computer readable media are directed to obtaining a point cloud that represents an environment based at least in part on a plurality of points in three-dimensional space; determining corresponding classifications of points in the point cloud as ground or not-ground based at least in part on a plurality of ground classification algorithms; determining respective point cloud features associated with the points in the point cloud; determining respective cell features associated with a plurality of cells that segment the point cloud; generating feature data for a machine learning model based at least in part on one or more of: the classifications of the points based on the plurality of ground classification algorithms, the point cloud features, or the cell features; and classifying the points in the point cloud based at least in part on an output from the machine learning model.
1. A computer-implemented method comprising:
obtaining a point cloud that represents an environment based at least in part on a plurality of points in three-dimensional space;
determining corresponding classifications of points in the point cloud as ground or not-ground based at least in part on a plurality of ground classification algorithms;
determining respective point cloud features associated with the points in the point cloud;
determining respective cell features associated with a plurality of cells that segment the point cloud;
generating feature data for a machine learning model based on at least two of: the classifications of the points based on the plurality of ground classification algorithms, the point cloud features, or the cell features; and
classifying the points in the point cloud based at least in part on an output from the machine learning model in response to input of the feature data.
2. The computer-implemented method of claim 1 , wherein determining the respective point cloud features comprises:
determining attributes for each point, the attributes including at least one of: coordinates of the point in three-dimensional space, a color associated with the point, an intensity associated with the point, or an estimated above ground level (AGL) associated with the point.
3. The computer-implemented method of claim 1 , wherein the output from the machine learning model further comprises respective ground classification probabilities for the points in the point cloud, wherein a ground classification probability for a point represents a likelihood of that point corresponding to a ground level in the environment.
4. The computer-implemented method of claim 3 , wherein the attributes determined for the cell comprise one or more of: a minimum z value of a point in the cell or a maximum z value of a point in the cell; a minimum, maximum, or median AGL associated with points included in the cell; a standard deviation of AGL associated with points included in the cell; a minimum, maximum, or median intensity of points included in the cell; a minimum, maximum, or median curvature associated with points included in the cell; a minimum, maximum, or median red, green, blue (RGB) values associated with points included in the cell; a minimum, maximum, or median normalized different vegetation index (NDVI); a minimum, maximum, median, or standard deviation of zenith angle (or slope) associated with points included in the cell; a minimum, maximum, or median simple morphological filter (SMRF) estimated AGL; a standard deviation of normals; a minimum, maximum, or median interquartile range (IQR) of vertical z coordinates, or a combination thereof.
5. The computer-implemented method of claim 1 , wherein the plurality of points in three-dimensional space is determined by a sensor device.
6. The computer-implemented method of claim 1 , wherein determining the respective cell features associated with the plurality of cells comprises:
dividing the point cloud into a two-dimensional grid of cells along horizontal X and Y dimensions, wherein a cell corresponds to a portion of the environment as represented by the point cloud; and
determining one or more attributes for each cell, wherein the one or more attributes are representative of points included in the cell.
7. The computer-implemented method of claim 1 , further comprising:
determining respective regional features associated with a plurality of regions of cells that represent the environment, wherein a region of cells comprises a plurality of cells.
8. The computer-implemented method of claim 7 , wherein determining the respective regional features associated with the plurality of regions of cells that represent the environment comprises:
grouping cells in the plurality of cells into the plurality of regions of cells; and
determining one or more attributes for each region of cells, wherein the one or more attributes are representative of a plurality of cells and points included in the region of cells.
9. The computer-implemented method of claim 8 , wherein the attributes determined for the region of cells comprise one or more of: a height in radius R associated with the region of cells; a depth in radius R associated with the region of cells; a Z entropy in radius R associated with the region of cells; a minimum height of cell below minimum height in some radius associated with the region of cells; an intensity entropy in the region of cells, a slope of a cell in the region of cells that is above a median slope of the point cloud, or a combination thereof.
10. The computer-implemented method of claim 1 , wherein generating the feature data for the machine learning model comprises:
formatting the feature data for ingestion by the machine learning model.
11. A system comprising at least one processor and memory storing instructions that cause the system to perform:
obtaining a point cloud that represents an environment based at least in part on a plurality of points in three-dimensional space;
determining corresponding classifications of points in the point cloud as ground or not-ground based at least in part on a plurality of ground classification algorithms;
determining respective point cloud features associated with the points in the point cloud;
determining respective cell features associated with a plurality of cells that segment the point cloud;
generating feature data for a machine learning model based at least on two of: the classifications of the points based on the plurality of ground classification algorithms, the point cloud features, or the cell features; and
classifying the points in the point cloud based at least in part on an output from the machine learning model in response to input of the feature data.
12. The system of claim 11 , wherein determining the respective point cloud features causes the system to perform:
determining attributes for each point, the attributes including at least one of: coordinates of the point in three-dimensional space, a color associated with the point, an intensity associated with the point, or an estimated above ground level (AGL) associated with the point.
13. The system of claim 11 , wherein the output from the machine learning model further comprises respective ground classification probabilities for the points in the point cloud, wherein a ground classification probability for a point represents a likelihood of that point corresponding to a ground level in the environment.
14. The system of claim 13 , wherein the plurality of points in three-dimensional space is determined by a sensor device.
15. The system of claim 11 , wherein determining the respective cell features associated with the plurality of cells causes the system to perform:
dividing the point cloud into a two-dimensional grid of cells along horizontal X and Y dimensions, wherein a cell corresponds to a portion of the environment as represented by the point cloud; and
determining one or more attributes for each cell, wherein the one or more attributes are representative of points included in the cell.
16. A system comprising at least one processor and memory storing instructions that cause the system to perform:
obtaining a point cloud that represents an environment based at least in part on a plurality of points in three-dimensional space;
determining corresponding classifications of points in the point cloud as ground or not-ground based at least in part on a plurality of ground classification algorithms;
determining respective point cloud features associated with the points in the point cloud;
determining respective cell features associated with a plurality of cells that segment the point cloud;
generating feature data for a machine learning model based on at least two of: the classifications of the points based on the plurality of ground classification algorithms, the point cloud features, or the cell features; and
training the machine learning model based at least in part on the feature data.
17. The system of claim 16 , wherein the plurality of points in three-dimensional space is determined by a sensor device.
18. The system of claim 16 , wherein the feature data further comprises ground truth information for the points in the point cloud, wherein the ground truth information correctly classifies the points in the point cloud as ground or not-ground.
19. The system of claim 18 , wherein the instructions further cause the at least one processor to perform:
storing the feature data as a training example in a training data repository.
20. The system of claim 16 , wherein the instructions further cause the at least one processor to perform:
determining respective regional features associated with a plurality of regions of cells that represent the environment, wherein a region of cells comprises a plurality of cells.