Object association for autonomous vehicles
Systems, methods, tangible non-transitory computer-readable media, and devices for associating objects are provided. For example, the disclosed technology can receive sensor data associated with the detection of objects over time. An association dataset can be generated and can include information associated with object detections of the objects at a most recent time interval and object tracks of the objects at time intervals in the past. A subset of the association dataset including the object detections that satisfy some association subset criteria can be determined. Association scores for the object detections in the subset of the association dataset can be determined. Further, the object detections can be associated with the object tracks based on the association scores for each of the object detections in the subset of the association dataset that satisfy some association criteria.
1. A computer-implemented method for machine-learned model training, the method comprising:
inputting, into an object association model, training data indicative of a plurality of training objects and a plurality of training object tracks in an environment over a plurality of time intervals;
receiving, from the object association model, an output indicative of a training association of at least one training object at a most recent time interval of the plurality of time intervals and at least one training object track at a plurality of time intervals preceding the most recent time interval;
determining, for the object association model, a similarity score based on a comparison of the training association to a ground truth association, wherein the similarity score is positively correlated with greater accuracy of association by the object association model; and
adjusting at least one parameter of the object association model based on the similarity score.
2. The computer-implemented method of claim 1 , wherein the training objects are indicative of at least one of: (i) a vehicle, (ii) a pedestrian, (iii) a road, or (iv) a structure.
3. The computer-implemented method of claim 1 , further comprising:
generating the similarity score based on respective classified object labels associated with the training objects.
4. The computer-implemented method of claim 1 , further comprising:
determining the similarity score is below a threshold similarity score indicating an inaccurate detection.
5. The computer-implemented method of claim 4 , wherein the threshold similarity score is based on similarity scores recorded in a plurality of previous training sessions of the object association model.
6. The computer-implemented method of claim 1 , further comprising:
determining an accuracy of association value based on the training association of the at least one training object, wherein the accuracy of association value is indicative of a portion of the training objects that are correctly associated by the object association model during training.
7. The computer-implemented method of claim 1 , further comprising:
determining an accuracy of tracking value based on the at least one training object track, wherein the accuracy of tracking value is indicative of a portion of the plurality of training object tracks that are correctly tracked by the object association model during training.
8. A computing system configured to train a machine-learned model comprising:
one or more processors; and
one or more non-transitory computer-readable media storing instructions executable by the one or more processors to perform operations, the operations comprising:
inputting, into an object association model, training data indicative of a plurality of training objects and a plurality of training object tracks in an environment over a plurality of time intervals;
receiving, from the object association model, an output indicative of a training association of at least one training object at a most recent time interval of the plurality of time intervals and at least one training object track at a plurality of time intervals preceding the most recent time interval;
determining, for the object association model, a similarity score based on a comparison of the training association to a ground truth association, wherein the similarity score is positively correlated with greater accuracy of association by the object association model; and
adjusting at least one parameter of the object association model based on the similarity score.
9. The computing system of claim 8 , wherein the training objects are indicative of at least one of: (i) a vehicle, (ii) a pedestrian, (iii) a road, or (iv) a structure.
10. The computing system of claim 8 , wherein the operations further comprise:
generating the similarity score based on respective classified object labels associated with the training objects.
11. The computing system of claim 8 , wherein the operations further comprise:
determining the similarity score is below a threshold similarity score indicating an inaccurate detection.
12. The computing system of claim 11 , wherein the threshold similarity score is based on similarity scores recorded in a plurality of previous training sessions of the object association model.
13. The computing system of claim 8 , further comprising:
determining an accuracy of association value based on the training association of the at least one training object, wherein the accuracy of association value is indicative of a portion of the training objects that are correctly associated by the object association model during training.
14. The computing system of claim 8 , further comprising:
determining an accuracy of tracking value based on at least one training object track, wherein the accuracy of tracking value is indicative of a portion of the plurality of training object tracks that are correctly tracked by the object association model during training.
15. A non-transitory computer-readable media storing instructions executable by one or more processors to cause the one or more processors to perform operations, the operations comprising:
inputting, into an object association model, training data indicative of a plurality of training objects and a plurality of training object tracks in an environment over a plurality of time intervals;
receiving, from the object association model, an output indicative of a training association of at least one training object at a most recent time interval of the plurality of time intervals and at least one training object track at a plurality of time intervals preceding the most recent time interval;
determining, for the object association model, a similarity score based on a comparison of the training association to a ground truth association, wherein the similarity score is positively correlated with greater accuracy of association by the object association model; and
adjusting at least one parameter of the object association model based on the similar score.
16. The non-transitory computer-readable media of claim 15 , wherein the training objects are indicative of at least one of: (i) a vehicle, (ii) a pedestrian, (iii) a road, or (iv) a structure.
17. The non-transitory computer-readable media of claim 15 , further comprising:
generating the similarity score based on respective classified object label associated with the training objects.
18. The non-transitory computer-readable media of claim 15 , further comprising:
determining the similarity score is below a threshold similarity score indicating an inaccurate detection.
19. The non-transitory computer-readable media of claim 18 , wherein the threshold similarity score is based on similarity scores recorded in a plurality of previous training sessions of the object association model.
20. The non-transitory computer-readable media of claim 15 , further comprising:
determining an accuracy of association value based on the training association of the at least one training object, wherein the accuracy of association value is indicative of a portion of the training objects that are correctly associated by the object association model during training.