Tracking across multiple cameras with disjoint views
First Claim
1. A method of tracking an object passing before non-overlapping cameras, comprising the steps of:
- tracking the object between a first camera and a second camera; and
automatically determining whether the object is identical in both the first camera and the second camera without calibrating the cameras or providing a site modeling, wherein the determining step includes the steps of;
learning inter-camera spatial temporal probability between the first camera and the second camera using Parzen windows;
learning inter-camera appearance probabilities between the first camera and the second camera using distribution of Bhattacharyya distances between appearance models;
establishing correspondences between the first camera and the second camera based on Maximum A Posteriori (MAP) framework combining both the spatial temporal and the appearance probabilities; and
updating learned probabilities throughout over time.
1 Assignment
0 Petitions
Accused Products
Abstract
Tracking and surveillance methods and systems for monitoring objects passing in front of non-overlapping cameras. Invention finds corresponding tracks from different cameras and works out which object passing in front of the camera(s) made the tracks, in order to track the object from camera to camera. The invention uses an algorithm to learn inter-camera spatial temporal probability using Parzen windows, learns inter-camera appearance probabilities using distribution of Bhattacharyya distances between appearance models, establishes correspondences based on Maximum A Posteriori (MAP) framework combining both spatial temporal and appearance probabilities, and updates learned probabilities throughout the lifetime of the system.
169 Citations
17 Claims
-
1. A method of tracking an object passing before non-overlapping cameras, comprising the steps of:
-
tracking the object between a first camera and a second camera; and automatically determining whether the object is identical in both the first camera and the second camera without calibrating the cameras or providing a site modeling, wherein the determining step includes the steps of; learning inter-camera spatial temporal probability between the first camera and the second camera using Parzen windows; learning inter-camera appearance probabilities between the first camera and the second camera using distribution of Bhattacharyya distances between appearance models; establishing correspondences between the first camera and the second camera based on Maximum A Posteriori (MAP) framework combining both the spatial temporal and the appearance probabilities; and updating learned probabilities throughout over time. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12)
-
-
13. A system for tracking an object passing between non-overlapping cameras without calibrating the cameras or completing site modeling, comprising:
-
plural cameras for tracking the object; and means for automatically determining whether the object in both a first one of the plural cameras and a second one of the plural cameras are a same object, wherein the determining means includes; means for learning inter-camera spatial temporal probability between a first camera and the second camera using Parzen windows, wherein the spatial temporal probability learning means includes means for collecting plural space-time features from the first camera and the second camera; and means for estimating a space-time probability density function for the first camera and the second camera; means for learning inter-camera appearance probabilities between the first camera and the second camera using distribution of Bhattacharyya distances between appearance models; means for establishing correspondences between the first camera and the second camera based on Maximum A Posteriora framework combining both the learned inter-camera spatial temporal and appearance probabilities; and means for updating the learned probabilities throughout over time. - View Dependent Claims (14, 15)
-
-
16. A system for tracking an object passing between non-overlapping cameras without calibrating the cameras or completing site modeling, comprising:
-
plural cameras for tracking the object; and means for automatically determining whether the object in both a first one of the plural cameras and a second one of the plural cameras are a same object, wherein the determining means includes; means for learning inter-camera spatial temporal probability between a first camera and the second camera using Parzen windows; means for learning inter-camera appearance probabilities between the first camera and the second camera using distribution of Bhattacharyya distances between appearance models wherein the appearance probability learning means includes; means for estimating a change of appearance of the object from the first camera to the second camera; means for establishing correspondences between the first camera and the second camera based on Maximum A Posteriora framework combining both the learned inter-camera spatial temporal and appearance probabilities; and means for updating the learned probabilities throughout over time. - View Dependent Claims (17)
-
Specification