Lane segment clustering using hybrid distance metrics
Disclosed herein are system, method, and computer program product embodiments for clustering lane segments of a roadway in order to improve and simplify autonomous vehicle behavior testing. The approaches disclosed herein provide a hybrid methodology of dividing lane segments into hard features and soft features, and using a metric learning model trained in a supervised process on the entirety of lane segment features to cluster the lane segments based on the soft features. These clustered lane segments can then be assigned to what is termed as protolanes, where a single set of tests applied to a given protolane is considered valid across all of the lane segments assigned to the protolane.
1. A method, comprising:
grouping, by one or more computing devices, a set of lane segments based on trained labels from a supervised learning process, wherein the trained labels are applicable to a plurality of features of the set of lane segments;
calculating, by the one or more computing devices, distances corresponding to the set of lane segments based on individual features of the set of lane segments assessed by a metric learning model trained with the trained labels;
clustering, by the one or more computing devices, the set of lane segments based on the distances; and
assigning, by the one or more computing devices, corresponding protolanes to clusters of the set of lane segments based on the clustering, wherein an additional lane segment would be assigned to one of the protolanes based on clustering the additional lane segment into a corresponding one of the clusters.
2. The method of claim 1 , further comprising:
presenting, by the one or more computing devices, two or more lane segments of the set of lane segments to a user;
determining, by the one or more computing devices, a similarity between the two or more lane segments of the set of lane segments based on a user input; and
assigning the trained labels to the two or more lane segments of the set of lane segments based on the similarity.
3. The method of claim 2 , wherein the similarity between the two or more lane segments of the set of lane segments based on the user input is based on a ranking supervision approach.
4. The method of claim 1 , further comprising:
binning, by the one or more computing devices, a plurality of lane segments into respective bins based on one or more hard features of the lane segments; and
selecting, by the one or more computing devices, the set of lane segments from a bin of the respective bins.
5. The method of claim 4 , further comprising:
extracting, by the one or more computing devices, individual features of the plurality of lane segments; and
routing, by the one or more computing devices, hard features of the individual features of the plurality of lane segments to execution of the binning, and soft features of the individual features of the plurality of lane segments to execution of the distance calculation.
6. The method of claim 1 , wherein the metric learning method utilizes a contrastive learning technique.
7. The method of claim 1 , further comprising:
determining, by the one or more computing devices, two lane segments associated with a protolane within a cluster of the clusters of the set of lane segments, such that the two lane segments are furthest apart from each other;
receiving, by the one or more computing devices, a determination that the two lane segments should be assigned to disparate protolanes; and
re-clustering, by the one or more computing devices, the set of lane segments based on the distances in an increased number of clusters sufficient to place the two lane segments into separate clusters.
8. A system, comprising:
a memory; and
at least one processor coupled to the memory and configured to execute operations comprising:
grouping a set of lane segments based on trained labels from a supervised learning process, wherein the trained labels are applicable to a plurality of features of the set of lane segments,
calculating distances corresponding to the set of lane segments based on individual features of the set of lane segments assessed by a metric learning model trained with the trained labels,
clustering the set of lane segments based on the distances, and
assigning corresponding protolanes to clusters of the set of lane segments based on the clustering, wherein an additional lane segment would be assigned to one of the protolanes based on clustering the additional lane segment into a corresponding one of the clusters.
9. The system of claim 8 , the operations further comprising:
presenting two or more lane segments of the set of lane segments to a user;
determining a similarity between the two or more lane segments of the set of lane segments based on a user input; and
assigning the trained labels to the two or more lane segments of the set of lane segments based on the similarity.
10. The system of claim 9 , wherein the similarity between the two or more lane segments of the set of lane segments based on the user input is based on a ranking supervision approach.
11. The system of claim 8 , the operations further comprising:
binning a plurality of lane segments into respective bins based on one or more hard features of the lane segments; and
selecting the set of lane segments from a bin of the respective bins.
12. The system of claim 11 , the operations further comprising:
extracting individual features of the plurality of lane segments; and
routing hard features of the individual features of the plurality of lane segments to execution of the binning, and soft features of the individual features of the plurality of lane segments to execution of the distance calculation.
13. The system of claim 8 , wherein the metric learning method utilizes a contrastive learning technique.
14. The system of claim 8 , the operations further comprising:
determining two lane segments associated with a protolane within a cluster of the clusters of the set of lane segments, such that the two lane segments are furthest apart from each other;
receiving a determination that the two lane segments should be assigned to disparate protolanes; and
re-clustering the set of lane segments based on the distances in an increased number of clusters sufficient to place the two lane segments into separate clusters.
15. A non-transitory computer-readable medium having instructions stored thereon that, when executed by at least one computing device, cause the at least one computing device to perform operations comprising:
grouping a set of lane segments based on trained labels from a supervised learning process, wherein the trained labels are applicable to a plurality of features of the set of lane segments;
calculating distances corresponding to the set of lane segments based on individual features of the set of lane segments assessed by a metric learning model trained with the trained labels;
clustering the set of lane segments based on the distances; and
assigning corresponding protolanes to clusters of the set of lane segments based on the clustering, wherein an additional lane segment would be assigned to one of the protolanes based on clustering the additional lane segment into a corresponding one of the clusters.
16. The non-transitory computer-readable medium of claim 15 , the operations further comprising:
presenting two or more lane segments of the set of lane segments to a user;
determining a similarity between the two or more lane segments of the set of lane segments based on a user input; and
assigning the trained labels to the two or more lane segments of the set of lane segments based on the similarity.
17. The non-transitory computer-readable medium of claim 16 , wherein the similarity between the two or more lane segments of the set of lane segments based on the user input is based on a ranking supervision approach.
18. The non-transitory computer-readable medium of claim 15 , the operations further comprising:
binning a plurality of lane segments into respective bins based on one or more hard features of the lane segments; and
selecting the set of lane segments from a bin of the respective bins.
19. The non-transitory computer-readable medium of claim 18 , the operations further comprising:
extracting individual features of the plurality of lane segments; and
routing hard features of the individual features of the plurality of lane segments to execution of the binning, and soft features of the individual features of the plurality of lane segments to execution of the distance calculation.
20. The non-transitory computer-readable medium of claim 15 , the operations further comprising:
determining two lane segments associated with a protolane within a cluster of the clusters of the set of lane segments, such that the two lane segments are furthest apart from each other;
receiving a determination that the two lane segments should be assigned to disparate protolanes; and
re-clustering the set of lane segments based on the distances in an increased number of clusters sufficient to place the two lane segments into separate clusters.