Object re-identification in video streams

Patent metadata

Field	Value
Publication number	US-12450864-B2
Application number	US-202318312126-A
Country	US
Kind code	B2
Filing date	May 4, 2023
Priority date	May 9, 2022
Publication date	Oct 21, 2025
Grant date	Oct 21, 2025

Abstract

The present disclosure generally relates to a method for weighting of features in a feature vector of an object detected in a video stream capturing a scene, comprising: determining a feature vector comprising a set of features for a detected object in the video stream; acquiring a reference feature vector of a reference model of the scene; and assigning a weight to at least one feature of the determined feature vector, wherein the weight for a feature of the determined feature vector depends on a deviation measure indicative of a degree of deviation of the feature from a corresponding feature of the acquired reference feature vector of the reference model.

First claim

The invention claimed is:

1. A method for performing re-identification of an object detected in at least one video stream capturing at least one scene, the method comprising:
determining, using a trained Convolutional Neural Network (CNN) algorithm for object detection and for determining feature vectors of detected objects, a first feature vector comprising a set of numerical features for a detected first object in a first image frame capturing one scene of the at least one scene;
determining, using the trained CNN algorithm, a second feature vector comprising a set of numerical features for a detected second object in a second image frame capturing the one scene of the at least one scene;
acquiring a reference feature vector calculated as an average or median feature vector of all or some of multiple feature vectors of a reference model of the one scene of the at least one scene, wherein the reference model is constructed to provide a representation of commonplace traits for the one scene of the at least one scene and is generated using object detections over a predetermined time period for generating an accurate model, and wherein the reference feature vector is pre-constructed prior to entering the determining steps;
assigning a weight to at least one numerical feature of the first feature vector, wherein the weight for a numerical feature of the first feature vector depends on a deviation measure indicative of a degree of deviation of the numerical feature of the first feature vector from a corresponding numerical feature of the acquired reference feature vector of the reference model, wherein numerical features of the first feature vector with larger deviations from the corresponding numerical feature of the reference feature vector are assigned higher weights than numerical features of the first feature vector with smaller deviations from the corresponding numerical feature of the reference feature vector;
assigning a weight to at least one numerical feature of the second feature vector, wherein the weight for a numerical feature of the second feature vector depends on a deviation measure indicative of a degree of deviation of the numerical feature of the second feature vector from a corresponding numerical feature of the reference feature vector of the reference model, wherein numerical features of the second feature vector with larger deviations from the corresponding numerical feature of the reference feature vector are assigned higher weights than numerical features of the second feature vector with smaller deviations from the corresponding numerical feature of the reference feature vector; and
re-identifying the detected second object as being the detected first object when the first feature vector is determined to correspond to the second feature vector according to a similarity measure between the first and second feature vectors being less than a threshold, the similarity measure is calculated using the assigned weights so that numerical features with higher weights are emphasized more than numerical features with lower weights in the calculation of the similarity measure.

2. The method according to claim 1 , wherein the method further comprises:
determining at least one further feature vector comprising a set of numerical features for the detected first object, wherein the step of assigning comprises:
assigning the weights to at least one numerical feature of the first feature vector further depending on a degree of similarity with a corresponding numerical feature of the determined at least one further feature vector.

3. The method according to claim 1 , wherein the at least one video stream is captured by a respective image acquisition device configured to monitor a single scene.

4. The method according to claim 1 , comprising assigning the same weights to numerical features of the first feature vector and to corresponding numerical features of the second feature vector, the weights being determined from an evaluation of the weights assigned to the first feature vector and the weights assigned to the second feature vector.

5. The method according to claim 1 , wherein the weights for the first feature vector of the detected first object is generated using a first reference model, and the weights for the second feature vector of the detected second object is generated using a second reference model.

6. The method according to claim 5 , wherein the weights assigned to a numerical feature of the first feature vector and to a corresponding numerical feature of the second feature vector depend on the numerical feature of the first feature vector and the numerical feature of the second feature vector that deviates most from the respective-acquired reference feature vector.

7. The method according to claim 5 , wherein the weights assigned to a numerical feature of the first feature vector and to a corresponding numerical feature of the second feature vector depend on the deviation from the reference feature vector of the numerical feature of the first feature vector and the numerical feature of the second feature vector that deviates the least from the respective reference feature vector.

8. The method according to claim 5 , comprising determining a first similarity measure using the first reference model, and a second similarity measure using the second reference model, wherein re-identifying comprises:
determining that the second object corresponds to the first object if any one of the first similarity measure and the second similarity measure is less than a threshold indicating that the first feature vector corresponds to the second feature vector.

9. An apparatus comprising a processor, the apparatus configured to execute the steps of a method for performing re-identification of an object detected in at least one video stream capturing at least one scene, the method comprising:
determining, using a trained Convolutional Neural Network (CNN) algorithm for object detection and for determining feature vectors of detected objects, a first feature vector comprising a set of numerical features for a detected first object in a first image frame capturing one scene of the at least one scene;
determining, using the trained CNN algorithm for object detection and for determining feature vectors of detected objects, a second feature vector comprising a set of numerical features for a detected second object in a second image frame capturing the one scene of the at least one scene;
acquiring a reference feature vector calculated as an average or median feature vector of all or some of multiple feature vectors of a reference model of a the one scene of the at least one scene, wherein the reference model is constructed to provide a representation of commonplace traits for the one scene of the at least one scene and is generated using object detections over a predetermined time period for generating an accurate model, and wherein the reference feature vector is pre-constructed prior to entering the determining steps;
assigning a weight to at least one numerical feature of the first feature vector, wherein the weight for a numerical feature of the first and second feature vector depends on a deviation measure indicative of a degree of deviation of the numerical feature of the first and second feature vector from a corresponding numerical feature of the reference feature vector of the reference model, wherein numerical features of the first feature vector with larger deviations from the corresponding numerical feature of the reference feature vector are assigned higher weights than numerical features of the first feature vector with smaller deviations from the corresponding numerical feature of the reference feature vector;
assigning a weight to

Assignees

• Axis Ab

Inventors

• Öhrn Anton
• Skans Markus
• Danielsson Niclas

Classifications

• G06V10/40

  Extraction of image or video features · CPC title

• G06V20/50 Primary

  Context or environment of the image · CPC title

• G06V10/443 Primary

  by matching or filtering · CPC title