A SEMI AUTOMATIC TARGET INITIALIZATION METHOD BASED ON VISUAL SALIENCY

• US 20160239981A1
• Filed: 08/28/2013
• Published: 08/18/2016
• Est. Priority Date: 08/28/2013
• Status: Active Grant

• Current Assignee: Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi
• Original Assignee: Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi
• Inventors: TUNALI, Emre, OZ, Sinan, AYTEKIN, Caglar
• Granted Patent: US 9,595,114 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06T 2207/20076 Probabilistic image processing
• CPC Class Codes: G06T 7/11 Region-based segmentation

Abstract

Target initialization can dramatically change the performance of the tracker, since the initial window determines for the tracker what to track. In order to achieve a better tracking performance; The present invention relates to a method of semi automatic target initialization based on visual saliency for a given point coordinate in the vicinity of target by the user. Performance boost of tracker is mainly based on two key features of target initialization algorithm: It is capable of compensating erroneous user input; also selecting the most distinctive, salient part of object as target, so better discrimination is achieved between the target and background. Experimental results show that tracking performance is boosted in scenarios, in which the tracking is initialized by the proposed algorithm. Very low computational cost and requirement of only a point coordinate as input in the neighborhood of the target make this approach preferable in real time tracking applications.

Claims

1. A method for semi automatic target initialization (100) fundamentally comprises the following steps,

image coordinate given by user (101),

initial window selection via CSD (102),

saliency map generation (103),

binarization (104)

target selection (105),

output, bounding box enclosing the target (106).

6. A method for semi automatic target initialization (100) according to claim 1, characterized as the step “

initial window selection via CSD (102)”

defines proper window that includes foreground and background which is capable of fulfilling the assumptions that are made in saliency generation method.

7. A method for semi automatic target initialization (100) according to claim 1, characterized as initial window selection which is started with histogram distance vector generation and finalized by selection of initial window size which corresponds 1^st local maximum of the histogram distance vector.

8. A method for semi automatic target initialization (100) according to claim 1, characterized as binarization which is achieved by calculating a thresholding saliency map where threshold is the weighted average of local maxima of saliency map with formula;

Thr=LocalMax_sorted^T·

∇

_norm²(LocalMax_sorted)

9. A method for semi automatic target initialization (100) according to claim 1, characterized as weights which are obtained by normalized laplacian of local maxima with formula;

10. A method for semi automatic target initialization (100) according to claim 1, characterized as target selection which is started with connected component analysis, then each component is inputted to the energy function and the component maximizing the regularization energy is selected as target and its bounding box is reported as the output of the system with the formula;

11. A method for semi automatic target initialization (100) according to claim 1, characterized by becoming semi-automatic and extremely fast by the giving click input in step “

image coordinate given by user (101)”

.

12. A method for semi automatic target initialization (100) according to claim 1, characterized as robust and user friendly since exactly same target initialization can be achieved even if the given coordinate scatters in a neighborhood.

13. A method for semi automatic target initialization (100) according to claim 1, characterized as suitable for long-term tracking since it boosts performance of tracker by achieving better discrimination between the target and background.

14. A method for semi automatic target initialization (100) according to claim 1 characterized by preventing error indicating at the starting of the saliency map by semi-automatic system in step “

image coordinate given by user (101)”

.

15. A method for semi automatic target initialization (100) according to claim 1 characterized by finding target bounding box with just a click by the user in step “

image coordinate given by user (101)”

turns the target initializer into a user friendly framework.

2. In the method for semi automatic target initialization (100), the step of “

initial window selection via CSD (102)”

further comprises the sub-steps of;

given coordinate input (201),

obtain foreground histogram (202),

obtain background histogram (203),

calculate quadratic-chi histogram distance (204),

Obtain histogram distance vector by repeating steps 202,203 and 204 for each window size (205),

select 1st local maximum of histogram distance vector as initial window size (206).

3. In the method for semi automatic target initialization (100), the step of “

saliency map generation (103)”

further comprises the sub-steps of;

given initial window (301),

calculate mean intensity for patches obtained by any super pixel selection algorithm (302),

calculate mean intensity difference in eight neighborhood and assign as weights (303),

calculate shortest path from each foreground patch to a background patch (304),

return saliency map by normalizing the values calculated as shortest path (305).

4. In the method for semi automatic target initialization (100), the step of “

binarization (104)”

further comprises the sub-steps of;

given saliency map (401),

exploit local maxima of saliency map and sort in descending order to obtain LocalMax_sorted (402),

calculate normalized laplacian of LocalMax_sorted (403),

obtain threshold by calculating weighted average of LocalMax_sorted (404),

binarize saliency map with threshold (405).

5. In the method for semi automatic target initialization (100), the step of “

target selection (105)”

further comprises the sub-steps of;

execute connected component analysis (501),

calculate regularization energy of each connected component (502),

report the bounding box of the connected component as target having the most salient region with minimum distance to center (503),