Systems and methods for utilizing machine learning models to reconstruct a vehicle accident scene from video

1. A method comprising:

receiving, by a device and from a vehicle device of a first vehicle, video data for video captured of a location associated with an accident associated with a second vehicle;

generating, by the device and based on the video data, a point cloud;

determining, by the device and based on the point cloud, an overhead view of the location associated with the accident;

projecting, by the device, bounding boxes of the second vehicle onto the overhead view to provide a trajectory of the second vehicle in the overhead view;

augmenting, by the device, the overhead view with the trajectory of the second vehicle, with additional data, to generate a final overhead view; and

performing, by the device, one or more actions based on the final overhead view.

2. The method of claim 1 , wherein performing the one or more actions comprises one or more of:

providing the final overhead view for display;

reporting the accident or poor driving behavior associated with the second vehicle based on the final overhead view;

scheduling a driver of the first vehicle for a defensive driving course based on the final overhead view;

causing the second vehicle to be disabled based on the final overhead view;

providing, to the second vehicle, a warning for a driver of the second vehicle based on the final overhead view; or

providing information about the second vehicle to an insurer of the second vehicle based on the final overhead view.

3. The method of claim 1 , wherein the point cloud is a dense semantic point cloud, and wherein generating, by the device and based on the video data, the dense semantic point cloud comprises:

processing, by the device, a dense point cloud and the video data, with a semantic segmentation model, to generate the dense semantic point cloud.

4. The method of claim 3 , wherein the dense point cloud is generated based on utilizing one or more depth maps with a sparse point cloud.

5. The method of claim 4 , wherein the one or more depth maps are depth maps for frames of the video data, and

wherein the depth maps are generated based on processing, by the device, the video data, with a convolutional neural network model, to generate the depth maps for the frames of the video data.

6. The method of claim 4 , wherein the sparse point cloud is associated with the location associated with the accident, and

wherein the sparse point cloud is generated based on processing, by the device, the video data, with a simultaneous localization and mapping model, to generate the dense semantic point cloud.

7. The method of claim 1 , wherein the overhead view is a semantic overhead view, and wherein determining the semantic overhead view comprises:

processing the point cloud, with a voxelization model, to determine the semantic overhead view.

8. A device, comprising:

one or more processors configured to:

receive, from a vehicle device of a first vehicle, video data for video captured of a location associated with an accident associated with a second vehicle;

process the video data, with a first model, to generate a semantic point cloud;

process the semantic point cloud, with a second model, to determine a semantic overhead view of the location associated with the accident;

project a plurality of images associated with second vehicle onto the semantic overhead view to provide a trajectory of the second vehicle in the semantic overhead view; and

augment the semantic overhead view, with the trajectory of the second vehicle, with additional data, to generate a final overhead view.

9. The device of claim 8 , wherein the one or more processors, to process the semantic point cloud, with the second model, to determine the semantic overhead view, are configured to:

quantize the semantic point cloud into voxels,

wherein each voxel includes points of the semantic point cloud associated with a respective region of the semantic point cloud; and

determine the semantic overhead view based on the voxels.

10. The device of claim 8 , wherein the plurality of images associated with the second vehicle comprise bounding boxes of the second vehicle, and

wherein the one or more processors, to project the plurality of images associated with the second vehicle onto the semantic overhead view to provide the trajectory of the second vehicle in the semantic overhead view, are configured to:

process the video data, with an object detection model, to determine the bounding boxes of the second vehicle;

process the bounding boxes, with a multiple object tracking model, to determine the trajectory of the second vehicle; and

include the trajectory of the second vehicle in the semantic overhead view.

11. The device of claim 8 , wherein the one or more processors, to augment the semantic overhead view with the trajectory of the second vehicle, with the additional data to generate the final overhead view, are configured to:

map satellite coordinates to the semantic overhead view; and

replace the semantic overhead view with an overhead satellite view based on the satellite coordinates to generate the final overhead view.

12. The device of claim 8 , wherein the additional data comprises one or more of:

one or more satellite images,

data identifying one or more road names,

data identifying one or more vehicle speeds, or

data identifying one or more vehicle distances.

13. The device of claim 8 , wherein the one or more processors, to augment the semantic overhead view with the trajectory of the second vehicle with the additional data to generate the final overhead view, are configured to:

calculate a speed of the second vehicle and one or more respective distances of the second vehicle from one or more other vehicles; and

present the speed and the one or more respective distances of the second vehicle in the final overhead view.

14. The device of claim 8 , wherein the video data includes video captured of the location by a plurality of cameras associated with the location.

15. A non-transitory computer-readable medium storing instructions, the instructions comprising:

one or more instructions that, when executed by one or more processors, cause the one or more processors to:

receive, from a vehicle device of a first vehicle, video data for video captured of a location associated with an accident associated with a second vehicle;

process the video data, with a segmentation model, to generate a point cloud;

process the point cloud, with a voxelization model, to determine an overhead view of the location associated with the accident;

project a plurality of images of the second vehicle onto the overhead view to provide a trajectory of the second vehicle in the overhead view; and

perform one or more actions based on the overhead view.

16. The non-transitory computer-readable medium of claim 15 , wherein the one or more instructions, that cause the one or more processors to perform the one or more actions, cause the one or more processors to one or more of:

provide the overhead view for display;

report the accident or poor driving behavior associated with the second vehicle based on the overhead view;

schedule a driver of the first vehicle for a defensive driving course based on the overhead view;

cause the second vehicle to be disabled based on the overhead view;

provide, to the second vehicle, a warning for a driver of the second vehicle based on the overhead view;

provide information about the second vehicle to an insurer of the second vehicle based on the overhead view; or

modify one or more of the segmentation model or the voxelization model based on the overhead view.

17. The non-transitory computer-readable medium of claim 15 , wherein the instructions further comprise:

one or more instructions that, when executed by the one or more processors, cause the one or more processors to:

augment the overhead view with the trajectory of the second vehicle, with additional data, to generate a final overhead view.

18. The non-transitory computer-readable medium of claim 15 , wherein the one or more instructions, that cause the one or more processors to process the video data, cause the one or more processors to:

utilize an object detection model to detect objects in frames of the video data; and

obtain, for the frames of the video data, respective bounding boxes and classes for the detected object s.

19. The non-transitory computer-readable medium of claim 15 , wherein the one or more instructions, that cause the one or more processors to process the video data, cause the one or more processors to:

generate depth maps for frames of the video data; and

generate, based on the depth maps and a sparse point cloud, a dense point cloud;

wherein the point cloud is generated based on the video data and the dense point cloud.

20. The non-transitory computer-readable medium of claim 15 , wherein the segmentation model is a semantic segmentation model and the point cloud is a semantic point cloud, and

wherein the one or more instructions, that cause the one or more processors to process the video data, with the semantic segmentation model, to generate the semantic point cloud, cause the one or more processors to:

assign semantic labels to pixels of frames of the video data; and

utilize corresponding semantic labels with a dense point cloud to generate the semantic point cloud.