Subspace projection based non-rigid object tracking with particle filters

• US 7,376,246 B2
• Filed: 06/27/2005
• Issued: 05/20/2008
• Est. Priority Date: 06/27/2005
• Status: Expired due to Fees

• Current Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Inventors: Shao, Jie, Porikli, Fatih M.
• Primary Examiner(s): Mehta; Bhavesh M.
• Assistant Examiner(s): SCHAFFER, JONATHAN C
• Application Number:
  US11/167,684
• Publication Number:
  US 20060291696A1
• Time in Patent Office: 1,058 Days
• Field of Search: 382/103
• US Class Current: 382/103
• CPC Class Codes: G06T 2207/30241 Trajectory
• CPC Class Codes: G06T 7/277 involving stochastic approaches, e.g. using Kalman filters
• CPC Class Codes: G06V 10/245 by locating a pattern; Special marks for positioning
• CPC Class Codes: G06V 10/755 Deformable models or variational models, e.g. snakes or active contours

Abstract

A method tracks non-rigid objects in a video acquired of a cluttered scene by a camera. The method uses a particle filter. The tracking includes the following steps: motion transition estimation, contour deformation detection, and contour regulation. The method uses a dynamic affine transform model and employs the particle filter to estimate the parameters of the model. The method generates a probabilistic map of deformation for tracking the contour of the object followed by a projection step to constrain or regulate the contour in a contour subspace.

Claims

1. A method for tracking a non-rigid object in a video, comprising for each frame in the video, the steps of:

estimating a motion transition contour model of the object;

assigning normal lines to the motion transition contour model;

constructing a posterior deformation probability map according to the normal lines;

detecting a deformed contour of the object according to the posterior deformation probability map;

obtaining a point distribution model for a generic contour-subspace;

projecting the deformed contour of the object onto a contour-subspace according to the point distribution model; and

regulating the projected contour.

2. The method of claim 1, in which the posterior deformation probability map includes a probabilistic contour template, a gradient magnitude map, a gradient orientation map, and a foreground probability map.

3. The method of claim 1, further comprising:

assigning, to pixels along the normal lines, a maximum probability of the posterior deformation probability map as the deformed contour.

4. The method of claim 1, further comprising:

selecting a set of curve support points for a current contour of the object;

assigning particles to the set of curve support points using a dynamic transition model;

constructing the normal lines on the support points;

extracting features for each normal line;

measuring a density of the features using a spatial Poisson distribution;

determining a likelihood score for each particle using the measured density;

updating a weight of each particle using a sampling proposal function;

determining a posterior probability function using the weights; and

assigning the maximum of the posterior probability as the motion transition contour model.

5. The method of claim 1, further comprising:

selecting a set of curve support points for a current contour of the object;

constructing the normal lines on the support points, and for each pixel along the normal lines, further comprising;

constructing a probabilistic contour template;

constructing a gradient magnitude map;

constructing a gradient orientation map; and

constructing a foreground probability map;

aggregating the probabilistic contour template, the gradient magnitude map, the gradient orientation map, and the foreground map to obtain the posterior deformation probability map; and

estimating a deformed position of each curve support point according to the posterior deformation probability map.

6. The method of claim 5, further comprising:

determining a pose variance score;

determining a length of each normal line;

generating a binary image from the motion transition model;

producing a distance map from the binary map; and

adapting parameters of the normal according to the distance map.

7. The method of claim 5, further comprising:

determining an edge map by diffusion, and constructing the gradient magnitude map and the gradient orientation map according to the edge map.

8. The method of claim 5, further comprising:

estimating a background image;

subtracting the frame from the background image; and

assigning a scaled difference as the foreground probability map.

9. The method of claim 5, further comprising:

determining a motion field for the frame;

estimating a global motion from the motion field;

subtracting the global motion from the motion field to obtain a difference; and

generating the foreground probability map from the difference.

10. The method of claim 1, further comprising:

estimating a point distribution map from a set of training samples; and

constructing the generic contour subspace using the point distribution model.