System and method of detecting objects

1. A computer implemented method for detecting an object in a scene comprising the steps of:

receiving, using a processor, classified threat patches having a depth map of 3D pixels of an image in the scene using a processor, wherein each said pixel includes associated 3D position data;

projecting, using said processor, said classified threat patches into a pre-generated vertical support histogram to facilitate selection of said projected threat patches having a score value within a sufficiency criterion, said vertical support histogram is a 2-D footprint of 3-D pixels in the image of the threat patches and the vertical support histogram is generated by retrieving and summing up height value of each of the 3-D pixel that satisfies depth and height constraints, according to:

N max row ( Z )= X res ×f x /Zx

N max col ( Z )= X res ×f x /Zx

N ( Z )= N max row ( Z )× N max col ( Z )

where N max row is maximum number of pixels in each image row, Z is distance of a rectangular grid cell from a camera, X res is a lateral value, f x is a horizontal focal length, N max col is maximum number of image rows in the height band [H max , H min ], f y is a vertical focal length, H max is the maximum height visible in the image at the distance Z and H min is the minimum height visible in the image at the distance Z;

grouping, using said processor, said selected patches in a plane fit to obtain a region of interest (ROI) in the image, said ROI comprising a plurality of grid cells with 2D pixels; and

processing, using said processor, said region of interest in the image to detect said object.

2. The method of claim 1 wherein said projecting comprising computing one of a mean, median or mean-around-median of the threat patches' (X,Z) coordinates to identify a region of interest (ROI) in the histogram, said ROI comprise a plurality of rectangular grid cells with 2D pixels.

3. The method of claim 2 further comprising computing a score value of sum of said cell s and matching the score value with the sufficiency criterion to select said projected threat patches.

4. The method of claim 1 wherein said grouping comprising the steps of:

(a) determining whether the selected threat patch is a candidate for a seed patch;

(b) identifying neighboring patches of the selected threat patch if said selected threat patch is the candidate seed for the patch;

(c) checking whether the candidate threat patch satisfy depth constraint;

(d) fitting the candidate threat patch in a plane of a current patch group if said candidate threat patch satisfies the depth constraints; said candidate threat patch is fitted using template planes with arbitrary orientations in the horizontal plane;

(e) determining whether each of the neighboring patches of the fitted candidate threat patch fit in the same plane of the current patch group; and

(f) adding each of the neighboring patches in the current patch group if the given neighboring patch fit a in the same plane of the current patch group;

(g) repeating steps (b) through (l) for all the selected threat patches that are the candidates for the seed patch.

5. The method of claim 4 further comprising repeating steps (b) through (f) for all the added neighboring patches in the current patch group.

6. The method of claim 4 further comprising determining whether the current patch group includes sufficient patches to create a group region of interest (ROI) around the object.

7. The method of claim 6 wherein if it is determined that the current group patch does not include sufficient patches, performing the steps of:

circumscribing the patches in the current group patch in a tessellated grid;

extending the patches according to their depth to maximize the ROI around the object;

selecting a center patch of the grid as the seed patch; and

repeating steps (b) through (l) of claim 4 .

8. The method of claim 6 further comprising creating a group ROI around the object if it is determined that the current group patch includes sufficient patches.

9. The method of claim 8 further comprising wherein said creating comprising the steps of:

computing maximum ROI bounding all the patches in the current group patch;

locating a set of number of top and bottom patches in the maximum ROI;

locating a set of number of left and right patches in the maximum ROI;

comparing the set of top and bottom matches with the set of left and right patches to retrieve the set with higher number of patches to form a refined Rot.

10. The method of claim 9 further comprising determining whether one or more portions of the refined ROI include in sufficient patches to detect said object.

11. The method of claim 10 further comprising identifying the one or more portions of the refined ROI including in sufficient patches.

12. The method of claim 11 further comprising splitting the refined ROI around the one or more portion s into one or more refined ROIs.

13. The method of claim 1 further comprising:

capturing imagery of the scene proximate a platform using a stereo image preprocessor;

producing the 3D depth map from the imagery using a depth map generator;

tessellating the depth map into a plurality of patches using a patch grid generator; and

classifying the plurality of patches as threat patches using the classifier module.

14. A system for detecting an object in a scene comprising:

a stereo image preprocessor module for capturing, digitizing and preprocessing a stereo imagery;

a depth map generator module for producing a 3D depth map from the imagery, wherein each pixel in the depth map includes associated 3D position data;

a vertical support generator module for processing said 3D depth maps to generate a vertical support histogram, said vertical support histogram is a 2-D footprint of a 3-D pixels in the depth map and the vertical support histogram is generated by retrieving and summing up height value of each of the 3-D pixel that satisfies depth and height constraints, according to:

N max row ( Z )= X res ×f x /Zx

N max col ( Z )= X res ×f x /Zx

N ( Z )= N max row ( Z )× N max col ( Z )

where N max row is maximum number of pixels in each image row, Z is distance of a rectangular grid cell from a camera, X res is a lateral value, f x is a horizontal focal length, N max col is a maximum number of image rows in the height band [H max , H min ], f y is a vertical focal length, H max is the maximum height visible in the image at the distance Z and H min is the minimum height visible in the image at the distance Z;

a patch grid generator for tessellating the depth map into a plurality of patches;

a classifier module for classifying said plurality of patches as threat patches;

a processor coupled to the classifier module and the vertical support generator module to function to project said classified threat patches into the vertical support histogram to facilitate selection of said projected threat patches having a score value within a sufficiency criterion, said processor further functions to group said selected threat patches having said score value using a plane fit to obtain a region of interest in the image and processes said region of interest in the image to detect said object.

15. A computer implemented method for detecting an object in a scene comprising the steps of:

receiving, using a processor, classified threat patches having a depth map of 3D pixels of an image in the scene using a processor, wherein each said pixel includes associated 3D position data;

projecting, using said processor, said classified threat patches into a pre-generated vertical support histogram to facilitate selection of said projected threat patches having a score value within a sufficiency criterion, said vertical support histogram is a 2-D footprint of 3-D pixels in the image of the threat patches;

grouping, using said processor, said selected patches in a plane fit to obtain a region of interest (ROI) in the image, said ROI comprising a plurality of grid cells with 2D pixels, the grouping comprising the steps of:

(a) determining whether the selected threat patch is a candidate for a seed patch;

(b) identifying neighboring patches of the selected threat patch if said selected threat patch is the candidate seed for the patch;

(c) checking whether the candidate threat patch satisfy depth constraint;

(d) fitting the candidate threat patch in a plane of a current patch group if said candidate threat patch satisfies the depth constraints; said candidate threat patch is fitted using template planes with arbitrary orientations in the horizontal plane;

(e) determining whether each of the neighboring patches of the fitted candidate threat patch fit in the same plane of the current patch group; and

(f) adding each of the neighboring patches in the current patch group if the given neighboring patch fit s in the same plane of the current patch group;

(g) repeating steps (b) through (l) for all the selected threat patches that are the candidates for the seed patch; and

processing, using said processor, said region of interest in the image to detect said object.