RAPID OBJECT LABELLING AND ANOMALY DETECTION FOR COMPUTER VISION AUTOMATIC TARGET RECOGNITION SYSTEMS
Methods, systems, and apparatuses, among other things, may label and classify objects via appearance-based clustering for computer vision automatic target recognition (ATR) systems, including automated anomaly detection for objects appearing in an area of interest (AOI).
1 . A method comprising:
receiving an area of interest;
determining, by a trained machine learning (ML) model, a feature vector associated with an object detected in the area of interest;
labeling the feature vector based on a characteristic associated with the object;
computing a Euclidean distance between the labeled feature vector and one or more stored feature vectors;
grouping a plurality of feature vectors into a cluster based on the computed Euclidean distance;
building a dendrogram of clusters depicting a hierarchy of the clusters;
labeling the object based on the dendrogram of clusters; and
presenting, by a user interface, a user with the labeled object.
2 . The method of claim 1 , wherein the area of interest is bounded by a polygon on a map.
3 . The method of claim 1 , wherein the area of interest is associated with an event.
4 . The method of claim 1 , wherein the feature vector comprises a bounding box associated with the object.
5 . The method of claim 1 , wherein the feature vector is determined based on feature extraction.
6 . The method of claim 1 , wherein the feature vector summarizes a visual appearance of the object.
7 . The method of claim 1 , wherein determining the feature vector comprises determining the object is not a part of a background associated with the area of interest.
8 . The method of claim 1 , wherein determining the feature vector comprises comparing an object to a previously detected object.
9 . The method of claim 1 , further comprising presenting, by a user interface, a user with the dendrogram of clusters.
10 . A method comprising:
determining, by a trained machine learning model, an anomaly score for a detected object;
determining an anomaly threshold;
determining an anomaly based on comparing the anomaly score to the threshold;
removing the anomaly from a set of detections;
determining an anomaly cluster comprising the anomaly; and
storing the anomaly cluster, wherein the anomaly score for the detected object is based on the anomaly cluster.
11 . The method of claim 10 , further comprising:
transmitting, to a user, the anomaly cluster; and
receiving, from the user, a confirmation or a rejection of the anomaly cluster, wherein the anomaly cluster is determined based on the confirmation or the rejection.
12 . The method of claim 10 , further comprising adding the determined anomaly to a set of anomalies.
13 . The method of claim 10 , wherein the anomaly is determined based on an anomaly detector.
14 . The method of claim 13 , further comprising training the anomaly detector.
15 . The method of claim 10 , wherein the anomaly threshold is determined based on an area of interest associated with the detected object.
16 . A computer program product comprising:
a computer-readable storage medium; and
instructions stored on the computer-readable storage medium that, when executed by a processor, causes the processor to:
receive an area of interest;
determine, by a trained machine learning (ML) model, a feature vector associated with an object detected in the area of interest;
label the feature vector based on a characteristic associated with the object;
compute a Euclidean distance between the labeled feature vector and one or more stored feature vectors;
group a plurality of feature vectors into a cluster based on the computed Euclidean distance;
build a dendrogram of clusters comprising the cluster and depicting a hierarchy of the clusters;
label the object based on the dendrogram of clusters; and
present, by a user interface, a user with the labeled object.
17 . The computer program product of claim 16 , wherein the feature vector comprises a bounding box associated with the object.
18 . The computer program product of claim 16 , wherein the feature vector summarizes a visual appearance of the object.
19 . The computer program product of claim 16 , wherein determining the feature vector comprises determining the object is not a part of a background associated with the area of interest.
20 . The computer program product of claim 16 , wherein determining the feature vector comprises comparing an object to a previously detected object.