Three-level scheme for efficient ball tracking

• US 8,184,855 B2
• Filed: 12/10/2007
• Issued: 05/22/2012
• Est. Priority Date: 12/10/2007
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Tong, Xiaofeng, Zhang, Yimin
• Primary Examiner(s): BHATNAGAR, ANAND P
• Application Number:
  US11/953,818
• Publication Number:
  US 20090147992A1
• Time in Patent Office: 1,625 Days
• Field of Search: 382/100, 382/103, 382/106, 382/107, 348/135, 348/143, 348/169, 348/170, 348/171, 348/172, 348/208.14
• US Class Current: 382/103
• CPC Class Codes: A63B 2024/0056 for statistical or strategic analysis
• CPC Class Codes: G06T 2207/10016 Video; Image sequence
• CPC Class Codes: G06T 2207/20061 Hough transform
• CPC Class Codes: G06T 2207/30224 Ball; Puck
• CPC Class Codes: G06T 2207/30241 Trajectory
• CPC Class Codes: G06T 7/251 involving models
• CPC Class Codes: G06V 40/20 Movements or behaviour, e.g. gesture recognition recognition of facial expressions G06V40/16

Abstract

A three-level ball detection and tracking method is disclosed. The ball detection and tracking method employs three levels to generate multiple ball candidates rather than a single one. The ball detection and tracking method constructs multiple trajectories using candidate linking, then uses optimization criteria to determine the best ball trajectory.

Claims

1. A ball detection and tracking method operable on a multiple-frame video image, the ball detection and tracking method being executable on a processor-based system, the ball detection and tracking method comprising:

receiving a global view frame of a frame of the multiple-frame video;

selecting multiple ball candidates from the global view frame;

generating initial trajectories through linking adjacent ball candidates in a spatial-temporal domain;

using path optimization to obtain a true trajectory from the initial trajectories obtained; and

using a Hough transform to remove lines that generate noise, wherein the noise impairs ball candidate selection.

2. The ball detection and tracking method of claim 1, further comprising:

receiving a frame of the multiple-frame video; and

sending the frame to the object level processing only if the frame is a global view frame.

3. The ball detection and tracking method of claim 2, further comprising:

discerning a dominant color in the frame by accumulating hue-saturation-value color histograms of the frame; and

segmenting the frame according to the dominant color.

4. A ball detection and tracking method operable on a multiple-frame video image presentable to a display on a processor-based system, the ball detection and tracking method comprising:

receiving a global view frame of a frame of the multiple-frame video;

selecting multiple ball candidates from the global view frame;

generating initial trajectories through linking adjacent ball candidates in a spatial-temporal domain;

using path optimization to obtain a true trajectory from the initial trajectories obtained; and

filtering one or more regions out of the global view frame, the filtered regions being unqualified for ball candidate selection based on size, area, eccentricity, and/or circularity factor.

5. The ball detection and tracking method of claim 4, obtaining multiple ball candidates from the global view frame further comprising:

using a support vector machine classifier with color information to verify ball candidates.

6. The ball detection and tracking method of claim 4, generating initial trajectories through linking adjacent ball candidates in a spatial-temporal domain further comprising:

using a motion estimation algorithm to generate the initial trajectories.

7. The ball detection and tracking method of claim 6, using a motion estimation algorithm to generate the initial trajectories further comprising:

for each ball candidate, using Kalman filtering with the formation;

X_k=AX<sub>k−

1</sub>+w_k

Z_k=HX_k+v_k

where X_k is a state vector at time k, Z_k is a measurement vector at time k, w_k is system noise, v_k is measurement noise, A is a state transition matrix and H is a measurement matrix.

8. The ball detection and tracking method of claim 7, further comprising:

for each ball candidate, setting

9. The ball detection and tracking method of claim 6, further comprising:

assigning a confidence value for each trajectory according to its length, the confidence value being either true, false, or pendent; and

removing those trajectories for which the confidence value is false.

10. The ball detection and tracking method of claim 4, using path optimization to obtain a true trajectory from the initial trajectories obtained further comprising:

defining a distance between two trajectories of the initial trajectories obtained; and

generating a distance graph based on the defined distance.

11. The ball detection and tracking method of claim 10, defining a distance between two trajectories of the initial trajectories obtained further comprising:

applying a time-line model to define a distance between the two trajectories.

12. The ball detection and tracking method of claim 10, further comprising:

using a Dijkstra algorithm to determine the shortest path between trajectories.

13. The ball detection and tracking method of claim 11, further comprising:

using cubic spline interpolation to link an interval between ends of the two trajectories.

14. The ball detection and tracking method of claim 4, selecting multiple ball candidates from the global view frame further comprising selecting four ball candidates from the global view frame.