Models for stop sign database creation
An improved system and method of determining presence of stop signs at intersection based on multiple types of data including ground truth data, telemetry data, and/or trip stills gathered from vehicle devices. The data may be analyzed and aggregated by one or more models and/or neural networks to develop a prediction regarding presence of stop signs at individual intersections.
1. A method performed by a feature identification system having one or more hardware computer processors and one or more non-transitory computer readable storage device storing software instructions executable by the feature identification system, the method comprising:
accessing map data including indications of intersections within a geographic region;
accessing vehicle data including a plurality of telemetry data points, each of the telemetry data points indicating at least a location and a speed of a vehicle, wherein the vehicle data is obtained from a plurality of vehicle device sensors;
associating telemetry data points with corresponding intersections within the geographic region;
for individual intersections:
generating a kernel density estimate (KDE) plot indicating speed on a first axis and distance from the intersection on a second axis, wherein the KDE plot is configured to indicate speed of vehicles as they approach the intersection;
accessing a neural network trained using positive image data of KDE plots known to be associated with stop signs and negative image data of KDE plots known to not be associated with stop signs, wherein the neural network is configured to output a probability of a stop sign at the individual intersection based on the KDE plot generated for the particular intersection; and
evaluating the neural network based on the generated KDE plot for the individual intersection to determine a telemetry probability for the individual intersection;
accessing vehicle data including a plurality of trip stills data points obtained from a plurality of vehicle device sensors, each of the trip stills data points indicating at least a trip stills image obtained from a forward facing camera and location information;
associating trip stills data points with corresponding intersections within the geographic region;
for individual trip stills data points:
accessing an object detection model trained to identify stop signs within images; and
evaluating the object detection model based on the trip stills image of the trip stills data point, to determine a trip stills probability of a stop sign in the trip stills image;
for individual intersections:
evaluating an aggregation model configured to output a combined prediction indicative of likelihood of a stop sign at the individual intersection, wherein inputs to the aggregation model include one or more of the telemetry probability for the individual intersection and one or more trip stills probability for trips stills data points associated with the individual intersection; and
updating map data with the combined prediction.
2. The method of claim 1 , wherein if both a telemetry probability and a predetermined quantity of trip stills probabilities are available for an individual intersection, the combined prediction is based on an aggregation of the telemetry probability and the trip stills probabilities.
3. The method of claim 1 , wherein the combined prediction indicates presence of a stop sign if the telemetry probability is above a first threshold and less than a minimum quantity of trip stills probabilities for the intersection are available.
4. The method of claim 3 , wherein the first threshold is 95%.
5. The method of claim 1 , wherein the combined prediction indicates presence of a stop sign if at least one trip stills probability is above a second threshold.
6. The method of claim 5 , wherein the second threshold is 90%.
7. The method of claim 1 , wherein the combined prediction indicates presence of a stop sign if at least two trip stills probabilities are above a third threshold.
8. The method of claim 7 , wherein the third threshold is 80%.
9. The method of claim 1 , further comprising, for each intersection:
evaluating each of a plurality of trip stills probabilities associated with the individual intersection; and
determining an overall trip stills probability for the individual intersection based on the plurality of trip stills probabilities.
10. The method of claim 9 , wherein the overall trip stills probability indicates a quantity of trip stills probabilities associated with the individual intersection.
11. The method of claim 1 , wherein a first individual intersection indicates a first direction of travel associated with two intersecting streets.
12. The method of claim 11 , wherein a second individual intersection indicates a second direction of travel associated with the two intersecting streets.
13. The method of claim 1 , wherein the aggregation model comprises a machine learning model trained on attributes of the telemetry model and trip stills model.
14. The method of claim 13 , wherein the attributes include one or more of vehicles at intersections, confidence in prediction of each model, location of stop sign in trip still.
15. A computing system comprising:
a hardware computer processor;
a non-transitory computer readable medium having software instructions stored thereon, the software instructions executable by the hardware computer processor to cause the computing system to perform operations comprising:
accessing map data including indications of intersections within a geographic region;
accessing vehicle data including a plurality of telemetry data points, each of the telemetry data points indicating at least a location and a speed of a vehicle, wherein the vehicle data is obtained from a plurality of vehicle device sensors;
associating telemetry data points with corresponding intersections within the geographic region;
for individual intersections:
generating a kernel density estimate (KDE) plot indicating speed on a first axis and distance from the intersection on a second axis, wherein the KDE plot is configured to indicate speed of vehicles as they approach the intersection;
accessing a neural network trained using positive image data of KDE plots known to be associated with stop signs and negative image data of KDE plots known to not be associated with stop signs, wherein the neural network is configured to output a probability of a stop sign at the individual intersection based on the KDE plot generated for the particular intersection; and
evaluating the neural network based on the generated KDE plot for the individual intersection to determine a telemetry probability for the individual intersection;
accessing vehicle data including a plurality of trip stills data points obtained from a plurality of vehicle device sensors, each of the trip stills data points indicating at least a trip stills image obtained from a forward facing camera and location information;
associating trip stills data points with corresponding intersections within the geographic region;
for individual trip stills data points:
accessing an object detection model trained to identify stop signs within images; and
evaluating the object detection model based on the trip stills image of the trip stills data point, to determine a trip stills probability of a stop sign in the trip stills image;
for individual intersections:
evaluating an aggregation model configured to output a combined prediction indicative of likelihood of a stop sign at the individual intersection, wherein inputs to the aggregation model include one or more of a telemetry probability for the individual intersection and one or more trip stills probability for trips stills data points associated with the individual intersection; and
updating map data with the combined prediction.
16. The computing system of claim 15 , wherein the operations further comprise, for each intersection:
evaluating each of a plurality of trip stills probabilities associated with the individual intersection; and
determining an overall trip stills probability for the individual intersection based on the plurality of trip stills probabilities.
17. The computing system of claim 16 , wherein the overall trip stills probability indicates a quantity of trip stills probabilities associated with the individual intersection.
18. The computing system of claim 15 , wherein a first individual intersection indicates a first direction of travel associated with two intersecting streets.
19. A non-transitory computer readable medium having software instructions stored thereon, the software instructions executable by a hardware computer processor to cause a computing system to perform operations comprising:
accessing map data including indications of intersections within a geographic region;
accessing vehicle data including a plurality of telemetry data points, each of the telemetry data points indicating at least a location and a speed of a vehicle, wherein the vehicle data is obtained from a plurality of vehicle device sensors;
associating telemetry data points with corresponding intersections within the geographic region;
for individual intersections:
generating a kernel density estimate (KDE) plot indicating speed on a first axis and distance from the intersection on a second axis, wherein the KDE plot is configured to indicate speed of vehicles as they approach the intersection;
accessing a neural network trained using positive image data of KDE plots known to be associated with stop signs and negative image data of KDE plots known to not be associated with stop signs, wherein the neural network is configured to output a probability of a stop sign at the individual intersection based on the KDE plot generated for the particular intersection; and
evaluating the neural network based on the generated KDE plot for the individual intersection to determine a telemetry probability for the individual intersection;
accessing vehicle data including a plurality of trip stills data points obtained from a plurality of vehicle device sensors, each of the trip stills data points indicating at least a trip stills image obtained from a forward facing camera and location information;
associating trip stills data points with corresponding intersections within the geographic region;
for individual trip stills data points:
accessing an object detection model trained to identify stop signs within images; and
evaluating the object detection model based on the trip stills image of the trip stills data point, to determine a trip stills probability of a stop sign in the trip stills image;
for individual intersections:
evaluating an aggregation model configured to output a combined prediction indicative of likelihood of a stop sign at the individual intersection, wherein inputs to the aggregation model include one or more of a telemetry probability for the individual intersection and one or more trip stills probability for trips stills data points associated with the individual intersection; and
updating map data with the combined prediction.
20. The non-transitory computer readable medium of claim 19 , wherein the aggregation model comprises a machine learning model trained on attributes of the telemetry model and trip stills model.