Discovering object pathways in a camera network

Patent metadata

Field	Value
Publication number	US-9532012-B1
Application number	US-201615191171-A
Country	US
Kind code	B1
Filing date	Jun 23, 2016
Priority date	Sep 28, 2015
Publication date	Dec 27, 2016
Grant date	Dec 27, 2016

Abstract

In an approach to tracking at least one target subject in a camera network, a search is started to find a target subject on a camera within a camera network. Features are extracted from the target subject and search queries are initiated in other nearby cameras within the camera network. Search queries attempt to detect target subjects and present the finds in a ranked order. Application of aggregate searches in multiple cameras and prior search results are used to improve matching results in the camera network; propagate a search of the target subject to discover the full pathway in the camera network; and project future occurrences of the target subject in subsequent cameras in the camera network.

First claim

What is claimed is:

1. A computer program product for tracking at least one target subject in a camera network, the computer program product comprising:
a non-transitory computer readable storage medium and program instructions stored on the non-transitory computer readable storage medium when executed by a computer processor to perform a method, comprising:
receiving, at a user interface, a selection of at least one target subject in a first image taken by a first camera from a user;
extracting, by a feature module, one or more features of the selected at least one target subject;
storing, by a feature database, the extracted one or more features of the selected at least one target subject;
correlating, by a correlation module, the extracted one or more features of the selected at least one target subject, among a set of at least two cameras;
applying, by the correlation module, a first equation which quantifies a probabilistic likelihood another target subject found in a second camera is the same as the target subject found in the first camera, wherein the probabilistic likelihood is defined in terms of L as:
L (Target Subject in First Camera→Target Subject in Second Camera)= F First Camera (Target Subject in First Camera) {circle around (x)}F Second Camera (Target Subject in Second Camera));

where {circle around (x)} is a correlation between the Target Subject in First Camera and Target Subject in Second Camera;
applying, by a feature weight module, a second equation which takes a linear combination of the correlated one or more extracted features to quantify a ranking while matching the other target subject found in the second camera and the target subject found in the first camera, and modify a criticality associated with a plurality of features among the extracted one or more features over another plurality of features among the extracted one or more features, wherein the ranking is described in terms of E as:
E =Sum i→n ( w c i w f i F i );

where E is the rank measure, Sum, is the linear summation of a feature, n is the number of correlation models, w c i is the correlation weight of the correlation using a feature i, w f i is the feature weight of the feature i for the current searched target subject, and Fi is the feature correlation model between the two cameras using a feature i;
ranking, by a ranking module, a plurality of potential matches for the first target based on the second equation, wherein a higher E value for a potential match among the plurality of potential matches is indicative of a higher rank based on parameters configured in the second equation;
automatically re-ranking, by the ranking module, the plurality of potential matches by reconfiguring the parameters in the first equation and the second equation;
iteratively performing, by the ranking module, one or more search processes with a new input in order to enhance a camera-to-camera correlation based on the one or more extracted features;
responsive to determining the selected at least one target subject is present in one or more images from a set of cameras, applying, by a predicting adjustment module, a first analytics solution, wherein the first analytics solution predicts an occurrence of the selected at least one target in a next camera based on finding the at least one selected target in a previous camera;
responsive to determining the selected at least one target subject is present in the one or more images from the set of cameras, applying, by the correlation refinement module, a second analytics solution, wherein the second analytics solution populates a first plurality of queries based on one or more correlation models; and
responsive to determining the selected at least one target subject is present in the one or more images from the set of cameras, applying, by the feature weight module, a third analytics solution, wherein the third analytics solution populates a second plurality of queries based on weighted criticality of the one or more extracted features.

Assignees

• IBM

Inventors

• Chen xiao long
• Lv Na
• Yan Zhe
• Zhai Yun
• Zhao Zhuo

Classifications

• G06F18/22

  Matching criteria, e.g. proximity measures · CPC title

• G06F16/5838

  using colour · CPC title

• G08B13/19645

  Multiple cameras, each having view on one of a plurality of scenes, e.g. multiple cameras for multi-room surveillance or for tracking an object by view hand-over · CPC title

• G06F17/30047

  Physics · mapped topic

• H04N7/181 Primary

  for receiving images from a plurality of remote sources · CPC title