Method and system for event detection by analysis of linear feature occlusion
The invention is a method for detecting events in an imaged scene by analyzing the occlusion of linear features in the background image. Linear features, curved or straight, in specific scene locations are either manually specified or automatically learned from an image or image sequence of the background scene. For each linear feature, an occlusion model determines whether the line or part of it is occluded. The locations of the lines of interest in the scene, together with their occlusion characterizations, collectively form a description of the scene for a particular image. An event, defined as a series of descriptions of the scene over an image sequence, can then be initially defined and subsequently detected automatically by the system. An example application of this is counting cars or people passing in front of a video camera.
1. A method for detecting scene events, comprising the steps of:
(a) capturing a sequence of background images using a camera;
(b) specifying a location of one or a plurality of background linear features, each background linear feature having edge points which are present for most of the images of the background image sequence;
(c) computing a state of occlusion of the background linear feature using a computer by;
(i) inputting an image from a second image sequence;
(ii) applying an edge detection method to the input image at the same location as the background linear feature to detect an input image linear feature; and
(iii) comparing the input image linear feature with the background linear feature;
such that the state of occlusion of the background linear feature is occluded if a portion of the background linear feature is not present in the input image from the second image sequence, and
the state of occlusion of the background linear feature is not occluded if the background linear feature is present in the input image from the second image sequence;
(d) creating a definition of a scene description as a plurality of background linear features in an image of said scene together with their state of occlusion;
(e) creating a definition of an event as a sequence of scene descriptions; and
(f) processing said image sequence to detect the occurrence of one or a plurality of said event.
2. The method according to claim 1 , wherein the method further comprises a step of specifying the location of at least a linear feature that comprises straight lines or curvilinear features,
wherein the linear feature is a group of edge points in a continuous and contiguous arrangement.
3. The method according to claim 1 , wherein the method further comprises a step of specifying the location of static linear features manually or computing the location of static linear features automatically.
4. The method according to claim 1 , wherein the method further comprises a step of specifying said scene description manually as a plurality of said static linear features with said static linear features' manually predefined said state of occlusion, or computing said scene description automatically from an image of said scene as a plurality of said static linear features with said static linear features' computed said state of occlusion.
5. The method according to claim 1 , wherein the method further comprises a step of defining said event manually as a sequence of predefined scene descriptions, or
learning said event automatically from computed said scene description of an exemplary image sequence of said event.
6. An apparatus for detecting scene events, comprising:
(a) means for capturing a sequence of background images using a camera;
(b) means for specifying a location of one or a plurality of background linear features, each background linear feature having edge points which are present for most of the images of the background image sequence;
(c) means for computing a state of occlusion of the background linear feature using a computer by;
(i) inputting an image from a second image sequence:
(ii) applying an edge detection method to the input image at the same location as the background linear feature to detect an input image linear feature; and
(iii) comparing the input image linear feature with the background linear feature;
such that the state of occlusion of the background linear feature is occluded if a portion of the background linear feature is not present in the input image from the second image sequence, and
the state of occlusion of the background linear feature is not occluded if the background linear feature is present in the input image from the second image sequence;
(d) means for creating a definition of a scene description as a plurality of background linear features in an image of said scene together with their state of occlusion;
(e) means for creating a definition of an event as a sequence of scene descriptions; and
(f) means for processing said image sequence to detect the occurrence of one or a plurality of said event.
7. The apparatus according to claim 6 , wherein the apparatus further comprises means for specifying the location of at least a linear feature that comprises straight lines or curvilinear features,
wherein the linear feature is a group of edge points in a continuous and contiguous arrangement.
8. The apparatus according to claim 6 , wherein the apparatus further comprises means for specifying the location of static linear features manually or means for computing the location of static linear features automatically.
9. The apparatus according to claim 6 , wherein the apparatus further comprises means for specifying said scene description manually as a plurality of said static linear features with said static linear features' manually predefined said state of occlusion, or
means for computing said scene description automatically from an image of said scene as a plurality of said static linear features with said static linear features' computed said state of occlusion.
10. The apparatus according to claim 6 , wherein the apparatus further comprises means for defining said event manually as a sequence of predefined scene descriptions, or
means for learning said event automatically from computed said scene description of an exemplary image sequence of said event.