Method, apparatus and computer program product for three dimensional feature extraction from a point cloud

1. An apparatus comprising at least one processor and at least one non-transitory memory including computer program code instructions, the computer program code instructions configured to, when executed, cause the apparatus to at least:

receive point cloud data representative of an environment captured by a LIDAR sensor;

compute voxel sequences from the point cloud data;

extract voxel-wise semantic features from the voxel sequences;

model voxel-wise temporal changes for each of a plurality of locations based on the voxel-wise semantic features using a convolutional long short-term memory network to generate the model with each of the plurality of locations weighted based on convolutional neural network features; and

classify objects in the environment as dynamic objects or static objects based on the modeled voxel-wise temporal changes and based on output of a spatial transformer network provided to a three-dimensional conditional random field layer to distinguish between LIDAR sensor movement and object movement to identify objects in the environment as dynamic objects or static objects.

2. The apparatus of claim 1 , wherein causing the apparatus to compute voxel sequences from the point cloud data comprises causing the apparatus to use a voxel cloud connectivity segmentation method to group voxels in point clouds into perceptually meaningful regions.

3. The apparatus of claim 1 , wherein causing the apparatus to extract voxel-wise semantic features from the voxel sequences comprises causing the apparatus to: process the voxel sequences through an encoder-decoder based network to extract the voxel-wise semantic features.

4. The apparatus of claim 1 , wherein the apparatus is further caused to:

generate a three-dimensional surface model of the environment including objects classified as static objects and excluding objects classified as dynamic objects.

5. The apparatus of claim 4 , wherein the apparatus is further caused to:

employ the three-dimensional surface model of the environment to facilitate autonomous vehicle control.

6. A computer program product comprising at least one non-transitory computer-readable storage medium having computer-executable program code instructions stored therein, the computer-executable program code instructions comprising program code instructions to:

receive point cloud data representative of an environment captured by a LIDAR sensor;

compute voxel sequences from the point cloud data;

extract voxel-wise semantic features from the voxel sequences;

model voxel-wise temporal changes for each of a plurality of locations based on the voxel-wise semantic features using a convolutional long short-term memory network to generate the model with each of the plurality of locations weighted based on convolutional neural network features; and

classify objects in the environment as dynamic objects or static objects based on the modeled voxel-wise temporal changes and based on output of a spatial transformer network provided to a three-dimensional conditional random field layer to distinguish between LIDAR sensor movement and object movement to identify objects in the environment as dynamic objects or static objects.

7. The computer program product of claim 6 , wherein the program code instructions to compute voxel sequences from the point cloud data comprises program code instructions to use a voxel cloud connectivity segmentation method to group voxels in point clouds into perceptually meaningful regions.

8. The computer program product of claim 6 , wherein the program code instructions to extract voxel-wise semantic features from the voxel sequences comprises program code instructions to: process the voxel sequences through an encoder-decoder based network to extract the voxel-wise semantic features.

9. The computer program product of claim 6 , further comprising program code instructions to:

generate a three-dimensional surface model of the environment including objects classified as static objects and excluding objects classified as dynamic objects.

10. The computer program product of claim 9 , further comprising program code instructions to:

employ the three-dimensional surface model of the environment to facilitate autonomous vehicle control.

11. A method comprising:

receiving point cloud data representative of an environment captured by a LIDAR sensor;

computing voxel sequences from the point cloud data;

extracting voxel-wise semantic features from the voxel sequences;

modeling voxel-wise temporal changes for each of a plurality of locations based on the voxel-wise semantic features using a convolutional long short-term memory network to generate the model with each of the plurality of locations weighted based on convolutional neural network features; and

classifying objects in the environment as dynamic objects or static objects based on the modeled voxel-wise temporal changes and based on output of a spatial transformer network provided to a three-dimensional conditional random field layer to distinguish between LIDAR sensor movement and object movement to identify objects in the environment as dynamic objects or static objects.

12. The method of claim 11 , wherein computing voxel sequences from the point cloud data comprises using a voxel cloud connectivity segmentation method to group voxels in point clouds into perceptually meaningful regions.

13. The method of claim 11 , wherein extracting voxel-wise semantic features from the voxel sequences comprises processing the voxel sequences through an encoder-decoder based network to extract the voxel-wise semantic features.

14. The method of claim 11 , further comprising:

generating a three-dimensional surface model of the environment including objects classified as static objects and excluding objects classified as dynamic objects; and

employing the three-dimensional surface model of the environment to facilitate autonomous vehicle control.