LOOP RETRAINING MACHINE-LEARNING MODELS FOR VEHICLE IDENTIFICATION
A system captures images of vehicles during a first tagging event and a second tagging event. The system applies a machine learning model to the images to determine identifications of vehicles, and determines a misidentification of a vehicle based on matching the identifications of the first tagging event and the second tagging event. The system generates correction data including a corrected identification of the vehicle, and generates additional training examples based on the corrected identification. The machine-learning model is then retrained with the additional training examples, and the retrained machine-learning model is then applied to identify vehicles from newly received images.
1 . A method for improving vehicle identification accuracy in a vehicle management system, comprising:
receiving images of vehicles during a first tagging event and a second tagging event at a managed facility;
determining identifications of vehicles by applying a machine-learning model to the images, the machine-learning model being trained over training examples including images of vehicles labeled with features associated with corresponding identifications of the vehicles;
determining a misidentification of a vehicle based on matching the identifications of the first tagging event and the second tagging event;
generating correction data including a corrected identification of the vehicle;
determining features associated with a corrected identification of the vehicle;
generating additional training examples by labeling the images of the vehicle with features associated with the corrected identification of the vehicle; and
retraining the machine-learning model with the additional training examples; and
applying the retrained machine-learning model to determine identifications of vehicles from newly received images.
2 . The method of claim 1 , wherein the first tagging event is an entry event during which a vehicle enters the managed facility or enters a zone of the managed facility, and the second tagging event is an exit event during which a vehicle exits the managed facility or exits the zone of the managed facility.
3 . The method of claim 2 , further comprising generating for display images of the vehicle associated with the misidentification to a client device of a user, wherein the correction data is received from the client device.
4 . The method of claim 3 , wherein determining the misidentification of the vehicle includes identifying a hanging exit event or a hanging entry event that cannot be matched to any entry event or exit event that shares a same identification of a vehicle.
5 . The method of claim 4 , further comprising receiving an indication from the client device, indicating matching a hanging exit event to a hanging entry event, and correcting identification of vehicle of at least one of the matched hanging exit event or hanging entry event.
6 . The method of claim 5 , wherein generating additional training examples includes labeling the images associated with the matched hanging exit event or hanging entry event with the corrected identification of the vehicle; and storing the labeled images as the additional training examples.
7 . The method of claim 4 , further comprising automatically matching a hanging exit event to a hanging entry event, wherein automatically matching the hanging exit event and the hanging entry event comprises:
comparing images associated with a plurality of hanging exit events with images associated with a plurality of hanging entry events to determine similarity scores between features of each pair of a hanging exit event and a hanging entry event;
identifying a pair of a hanging exit event and a hanging entry event that have a similarity score greater than a predetermined threshold;
matching the pair of hanging exit event and hanging entry event; and
correcting identification of vehicle associated with one event in the matched pair of the hanging exit event and hanging entry event to a remaining event in the pair.
8 . The method of claim 7 , wherein each identification of a license plate by applying the machine-learning model is associated with a confidence score indicating a likelihood that the identification is correct, and correcting identification of vehicle associated with one event in the pair of hanging exit event and hanging entry event comprises:
accessing a confidence score associated with the identification of the hanging exit event and a confidence score associated with the identification of the hanging entry event to identify a lower confidence score; and
updating the identification of vehicle associated with the lower confidence score to the identification of vehicle associated with a higher confidence score.
9 . The method of claim 7 , wherein generating additional training examples includes:
labeling images associated with features associated with the corrected identification of vehicle; and
storing the labeled images as the additional training examples.
10 . The method of claim 1 , wherein the machine-learning model includes a license plate identification model configured to identify a license plate of a vehicle, the license plate identification model being trained over images of license plates labeled with corresponding identifications of the license plates; and
retraining the machine-learning model includes retraining the license plate identification model.
11 . The method of claim 10 , wherein the license plate identification model includes a jurisdiction classification model configured to determine a jurisdiction of a license plate of a vehicle, the jurisdiction classification model trained over images of license plates labeled with corresponding jurisdictions of the license plates; and
retraining the license plate identification model includes retraining the jurisdiction classification model.
12 . The method of claim 1 , wherein the machine-learning model includes a vehicle make-and-model identification model configured to identify a make and model of a vehicle, the vehicle make-and-model identification model trained over images of vehicles labeled with corresponding make and model of the vehicles; and
retraining the machine-learning model includes retraining the vehicle model identification model.
13 . A non-transitory computer-readable medium comprising memory with instructions encoded thereon, the instructions comprising instructions to cause one or more processors to:
receive images of vehicles during a first tagging event and a second tagging event at a managed facility;
determine identifications of vehicles by applying a machine-learning model to the images, the machine-learning model being trained over training examples including images of vehicles labeled with features associated with identification of the vehicles;
determine a misidentification of a vehicle based on matching the identifications of the first tagging event and the second tagging event;
generate correction data including a corrected identification of the vehicle;
determine features associated with a corrected identification of the vehicle;
generate additional training examples by labeling the images of the vehicle with features associated with the corrected identification of the vehicle; and
retrain the machine-learning model with the additional training examples; and
apply the retrained machine-learning model to determine identifications of vehicles from newly received images.
14 . The non-transitory computer-readable medium of claim 13 , wherein the first tagging event is an entry event during which a vehicle enters the managed facility or enters a zone of the managed facility, and the second tagging event is an exit event during which a vehicle exits the managed facility or exits the zone of the managed facility.
15 . The non-transitory computer-readable medium of claim 14 , wherein the instructions further cause the one or more processors to generate for display images of the vehicle associated with the misidentification to a client device of a user, wherein the correction data is received from the client device.
16 . The non-transitory computer-readable medium of claim 15 , wherein determining the misidentification of the vehicle includes identifying a hanging exit event or a hanging entry event that cannot be matched to any entry event or exit event that shares a same identification of a vehicle.
17 . The non-transitory computer-readable medium of claim 16 , wherein the instructions further cause the one or more processors to receive an indication from the client device, indicating matching a hanging exit event to a hanging entry event, and correcting identification of license plate of at least one of the matched hanging exit event or hanging entry event.
18 . The non-transitory computer-readable medium of claim 17 , wherein generating additional training examples includes labeling the images associated with the matched hanging exit event or hanging entry event with the corrected identification of the vehicle; and storing the labeled images as the additional training examples.
19 . The non-transitory computer-readable medium of claim 16 , the instructions further cause the one or more processors to automatically match a hanging exit event to a hanging entry event, wherein automatically matching the hanging exit event and the hanging entry event comprises:
comparing images associated with a plurality of hanging exit events with images associated with a plurality of hanging entry events to determine similarity scores between features of each pair of a hanging exit event and a hanging entry event;
identifying a pair of a hanging exit event and a hanging entry event that have a similarity score greater than a predetermined threshold;
matching the pair of hanging exit event and hanging entry event; and
correcting identification of vehicle associated with one event in the matched pair of the hanging exit event and hanging entry event to a remaining event in the pair.
20 . A system comprising:
memory with instructions encoded thereon; and
one or more processors that, when executing the instructions, are caused to perform operations comprising:
receive images of vehicles during a first tagging event and a second tagging event at a managed facility;
determine identifications of vehicles by applying a machine-learning model to the images, the machine-learning model being trained over training examples including images of vehicles labeled with features associated with identification of the vehicles;
determine a misidentification of a vehicle based on matching the identifications of the first tagging event and the second tagging event;
generate correction data including a corrected identification of the vehicle;
determine features associated with a corrected identification of the vehicle;
generate additional training examples by labeling the images of the vehicle with features associated with the corrected identification of the vehicle; and
retrain the machine-learning model with the additional training examples; and
apply the retrained machine-learning model to determine identifications of vehicles from newly received images.