Method and system for finding correspondence between face camera views and behavior camera views
View Patent ↗The present invention is a method and system to provide correspondences between a face camera track and a behavior camera track, for the purpose of making correspondence between the data obtained from each track. First, multiple learning machines are trained so that each of the machines processes pairwise person images from a specific pose region, and estimates the likelihood of two person images belonging to the same person based on image appearances. Then, the system acquires a person image associated with a behavior camera track, determines the pose of the person image based on its floor position, and corrects the pose of the person image. The system also acquires person images from face camera images associated with a face camera track, and combines the images with corrected person images from the previous step to form pairwise person images. The pairwise person image is fed to the trained pose-dependent pairwise person verification machines according to the pose of the person images, to compute the appearance match scores between the pair of person images. Finally, the combination of the appearance match scores and the spatiotemporal match scores of the pair of person images determines whether or not the person images belong to the same person.
1. A method for making a correspondence between a face camera track and a behavior camera track, comprising the following steps of:
a) constructing a database of pairwise person images, where each pairwise person image comprises a person image from a face camera and a person image from a behavior camera,
b) annotating the database of pairwise person images based on whether the pair of person images in each pairwise person image belongs to the same person,
c) training at least a pairwise person verification machine to estimate appearance match scores between the pair of person images in the pairwise person images from the database,
d) acquiring a person image from the face camera track and a person image from the behavior camera track to form a second pairwise person image, and
e) estimating an appearance match score between the pair of person images in the second pairwise person image, using the pairwise person verification machines on the second pairwise person image to calculate correspondence between the face camera track and the behavior camera track.
2. The method according to claim 1 , wherein the method further comprises a step of sampling the pairwise person images for training, so that the number of matching pairs and the number of non-matching pairs are balanced.
3. The method according to claim 1 , wherein the method further comprises a step of estimating and correcting the pose of the person image based on the person's floor position relative to the camera axis.
4. The method according to claim 1 , wherein the method further comprises a step of dividing the view of the face camera or the behavior camera into regions to process person images within each region separately,
wherein the person image in each region goes through a similar degree of distortion due to the orientation of the camera with respect to the region.
5. The method according to claim 4 , wherein the method further comprises a step of training multiple pairwise person verification machines, wherein each pairwise person verification machine is specialized to a specific range of person image pose and distortion.
6. The method according to claim 1 , wherein the method further comprises a step of estimating the position of the face camera person image and the position of the behavior camera person image in a common coordinate system to measure the spatial proximity of the person images.
7. The method according to claim 1 , wherein the method further comprises a step of utilizing the spatiotemporal proximity of the face camera person image and the position of the behavior camera person image.
8. The method according to claim 7 , wherein the method further comprises a step of combining the appearance match score and the spatiotemporal proximity between the face camera person image and the position of the behavior camera person image to find correspondence between the face camera track and the behavior camera track.
9. The method according to claim 1 , wherein the method further comprises a step of acquiring multiple person images from the face camera track and multiple person images from the behavior camera track to process multiple pairwise person images for finding the best match.
10. The method according to claim 1 , wherein the method further comprises a step of making correspondences between face classification data obtained from the face camera track and behavior data obtained from the behavior camera track.
11. The method according to claim 10 , wherein the method further comprises a step of using demographics classification data or person identification data as the face classification data.
12. The method according to claim 10 , wherein the method further comprises a step of using shopper behavior data as the behavior data.
13. An apparatus for making a correspondence between a face camera track and a behavior camera track, comprising:
a) means for constructing a database of pairwise person images, where each pairwise person image comprises a person image from a face camera and a person image from a behavior camera,
b) means for annotating the database of pairwise person images based on whether the pair of person images in each pairwise person image belongs to the same person,
c) means for training at least a pairwise person verification machine to estimate appearance match scores between the pair of person images in the pairwise person images from the database,
d) means for acquiring a person image from the face camera track and a person image from the behavior camera track to form a second pairwise person image, and
e) means for estimating an appearance match score between the pair of person images in the second pairwise person image, using the pairwise person verification machines on the second pairwise person image to calculate correspondence between the face camera track and the behavior camera track.
14. The apparatus according to claim 13 , wherein the apparatus further comprises means for sampling the pairwise person images for training, so that the number of matching pairs and the number of non-matching pairs are balanced.
15. The apparatus according to claim 13 , wherein the apparatus further comprises means for estimating and correcting the pose of the person image based on the person's floor position relative to the camera axis.
16. The apparatus according to claim 13 , wherein the apparatus further comprises means for dividing the view of the face camera or the behavior camera into regions to process person images within each region separately,
wherein the person image in each region goes through a similar degree of distortion due to the orientation of the camera with respect to the region.
17. The apparatus according to claim 16 , wherein the apparatus further comprises means for training multiple pairwise person verification machines, wherein each pairwise person verification machine is specialized to a specific range of person image pose and distortion.
18. The apparatus according to claim 13 , wherein the apparatus further comprises means for estimating the position of the face camera person image and the position of the behavior camera person image in a common coordinate system to measure the spatial proximity of the person images.
19. The apparatus according to claim 13 , wherein the apparatus further comprises means for utilizing the spatiotemporal proximity of the face camera person image and the position of the behavior camera person image.
20. The apparatus according to claim 19 , wherein the apparatus further comprises means for combining the appearance match score and the spatiotemporal proximity between the face camera person image and the position of the behavior camera person image to find correspondence between the face camera track and the behavior camera track.