Systems and methods for near-crash determination

1. A method for near-collision analysis, comprising, at an on-board system mounted to a vehicle:

recording an external video with an external-facing camera of the on-board system;

determining obstacle parameters for an obstacle detected in the external video;

generating a spatial risk map for the vehicle based on the obstacle parameters, the spatial risk map comprising a risk score for each of a plurality of spatial positions relative to the vehicle wherein the risk score is determined using:

a parametric model comprising a set of Gaussian models;

the obstacle parameters; and

driver parameters extracted from an internal video recorded by an internal-facing camera of the on-board system;

detecting a near-collision event when a risk score within the spatial risk map exceeds a risk threshold;

labeling the near-collision event with a label determined based on the driver parameters;

determining a cause of the near-collision event, comprising identifying an independent parameter of the parametric model with a highest weighted value; and

transmitting the label to a remote computing system.

2. The method of claim 1 , further comprising determining secondary obstacle parameters for a second obstacle detected in the external video, wherein each risk score is calculated based on the obstacle parameters and the secondary obstacle parameters.

3. The method of claim 1 , wherein the cause of the near-collision event is determined based on the driver parameters, wherein the label comprises the driver cause.

4. The method of claim 1 , further comprising transmitting a segment of the external video and a segment of the internal video to the remote computing system, in association with the label.

5. The method of claim 4 , further comprising:

determining a driver response to the near-collision event from the internal video;

aggregating driver responses to near-collision events for a plurality of drivers;

identifying good drivers associated with high driver scores;

determining the driver responses of the good drivers to near-collision events sharing a common label; and

training a vehicle control system based on the driver responses of the good drivers, the associated segment of the external video, and the associated segment of the internal video, wherein the vehicle control system controls an autonomous vehicle.

6. A method for near-collision analysis, comprising:

sampling an external image with an external-facing camera mounted to a vehicle;

determining obstacle parameters for an obstacle detected from the external image;

generating a spatial risk map for the vehicle based on the obstacle parameters, the spatial risk map comprising a risk score for each of a set of spatial positions relative to the vehicle, wherein the risk score is determined using:

a parametric model comprising a set of Gaussian models;

the obstacle parameters; and

driver parameters associated with a driver of the vehicle;

detecting a near-collision event when a risk score within the spatial risk map exceeds a risk threshold;

labeling the near-collision event with a label determined based on the driver parameters;

transmitting the label to a remote computing system; and

determining a cause of the near-collision event, comprising identifying an independent parameter of the parametric model with a highest weighted value.

7. The method of claim 6 , wherein a processing system on-board the vehicle detects and labels the near-collision event.

8. The method of claim 6 , further comprising determining secondary obstacle parameters for a second obstacle detected from the external image, wherein the spatial risk map is further generated based on the secondary obstacle parameters.

9. The method of claim 6 , wherein the driver parameters comprise a driver response determined from sensor signals sampled during the near-collision event.

10. The method of claim 9 , wherein the sensor signals comprise an internal image sampled by an internal-facing camera mounted to the vehicle.

11. The method of claim 9 , further comprising:

aggregating driver responses to near-collision events for a plurality of drivers;

identifying good drivers associated with high driver scores;

determining the driver responses of the good drivers; and

training a vehicle control system based on the driver responses of the good drivers, wherein the vehicle control system controls an autonomous vehicle.

12. The method of claim 11 , further comprising:

associating the near-collision event with a driver of the plurality; and

for each of the plurality of drivers, determining a driver score based on a history of near-collision events associated with the driver.

13. The method of claim 12 , further comprising coaching the driver based on the history of near-collision events associated with the driver.

14. The method of claim 6 , wherein, in response to the independent parameter with the highest weighted value comprising a driver parameter, determining a driver cause based on an internal image sampled by an internal-facing camera before the near-collision event, wherein the near-collision event is labeled with the driver cause.

15. The method of claim 14 , wherein the internal-facing camera is statically mounted at a known orientation to the external-facing camera, wherein determining the driver cause comprises determining a driver gaze direction relative to the obstacle based on the known orientation.

16. The method of claim 6 , further comprising:

determining a near-collision location associated with the near-collision event; and

aggregating near-collision locations of near-collision events from a plurality of vehicles into a geographic risk map.