Side window detection in near-infrared images utilizing machine learning

1. A side window detection method, comprising:

generating a deformable part model with respect to an image of a vehicle captured in a near-infrared band utilizing a side window detection and a B-frame detection module in order to obtain a set of candidate side-windows;

generating a refined deformable part model utilizing super pixel generation and a longest-line detection in order to remove a false alarm with respect to said deformable part model; and

refining detection performance of said refined deformable part model based on global regional information utilizing a local self-similarity based metric.

2. The method of claim 1 further comprising identifying a spatial feature associated with presence of an occupant by said side window detection and B-frame detection module.

3. The method of claim 1 further comprising:

encoding an elastic deformation and 3D structure of an object for face detection and pose estimation;

modeling a topological change due to a view point utilizing a mixture of trees with a shared pool of parts defined at each landmark position and a global mixture wherein said global mixture captures a gross deformation change for a single viewpoint; and

considering said refined deformable part model as a linear classifier, which generates a face detection score for an image region and assigning a highest score as a face region.

4. The method of claim 1 further comprising globally optimizing said side window detection with dynamic programming and still capture to detect a backseat side window boundary utilizing a B-pillar.

5. The method of claim 1 wherein said longest line detection further comprises:

computing an image derivative followed by a non-maximum suppression utilizing a canny edge detector and filling a line on a gradient image;

quantizing a gradient direction into a set of ranges, where all pixels having an orientation within a certain range fall into a corresponding bin and are assigned a particular label; and

connecting an edge pixel having a similar label via a connected component analysis algorithm with length above a certain threshold.

6. The method of claim 5 further comprising:

summing a binary line image in X direction and performing a low pass filtering of one dimensional line sum image; and

employing a longest line in said image of said vehicle that corresponds to a upper/lower side of a side window for potential side-window detection.

7. The method of claim 1 wherein said super pixel generation further comprises:

deriving a superpixel by applying a watershed transform to a grayscale side view image wherein said superpixel with a highest number of elements below a longest line typically corresponds to a body of said vehicle.

8. The method of claim 7 further comprising wherein said body of said vehicle is not considered for said side-window detection unless face detection score is high in order to reduce a false alarm rate.

9. The method of claim 1 wherein local self similarity processing further comprises:

adding a weighting term to a face detection score in order to remove a homogenous false alarm region;

measuring a local self similarity at every landmark point by correlating a local patch centered at landmark with a larger surrounding image region by said weighting term; and

computing a sum of chi-square distance between a histogram of a local and larger region and incorporating said weighting term into said deformable part model and employing a post-processing step utilizing a detected landmark point.

10. The method of claim 9 wherein said local self similarity processing further eliminates a false alarm due to a strong edge point.

11. A side window detection system, comprising:

a processor;

a data bus coupled to said processor; and

a computer-usable medium embodying computer program code, said computer-usable medium being coupled to said data bus, said computer program code comprising instructions executable by said processor and configured for:

generating a deformable part model with respect to an image of a vehicle captured in a near-infrared band utilizing a side window detection and a B-frame detection module in order to obtain a set of candidate side-windows;

generating a refined deformable part model utilizing superpixel generation and a longest-line detection in order to remove a false alarm with respect to said deformable part model; and

refining detection performance of said refined deformable part model based on global regional information utilizing a local self-similarity based metric.

12. The system of claim 11 wherein said instructions are further configured for identifying a spatial feature associated with presence of an occupant by said side window detection and B-frame detection module.

13. The system of claim 11 wherein said instructions are further configured for:

encoding an elastic deformation and 3D structure of an object for face detection and pose estimation;

modeling a topological change due to a view point utilizing a mixture of trees with a shared pool of parts defined at each landmark position and a global mixture wherein said global mixture captures a gross deformation change for a single viewpoint; and

considering said refined deformable part model as a linear classifier, which generates a face detection score for an image region and assigning a highest score as a face region.

14. The system of claim 11 wherein said instructions are further configured for globally optimizing said side window detection with dynamic programming and still capture to detect a backseat side window boundary utilizing a B-pillar.

15. The system of claim 11 wherein said instructions are further configured for:

computing an image derivative followed by a non-maximum suppression utilizing a canny edge detector and fitting a line on a gradient image;

quantizing a gradient direction into a set of ranges; where all pixels having an orientation within a certain range fall into a corresponding bin and are assigned a particular label; and

connecting an edge pixel having a similar label via a connected component analysis algorithm with length above a certain threshold.

16. The system of claim 15 wherein said instructions are further configured for:

summing a binary line image in X direction and performing a low pass filtering of one dimensional line sum image; and

employing a longest line in said image of said vehicle that corresponds to a upper/lower side of a side window for potential side-window detection.

17. The system of claim 11 wherein said superpixel generation further comprises:

deriving a superpixel by applying a watershed transform to a grayscale side view image wherein said superpixel with a highest number of elements below a longest line typically corresponds to a body of said vehicle.

18. The system of claim 11 wherein said instructions are further configured for:

adding a weighting term to a face detection score in order to remove a homogenous false alarm region;

measuring a local self similarity at every landmark point by correlating a local patch centered at landmark with a larger surrounding image region by said weighting term; and

computing a sum of chi-square distance between a histogram of a local and larger region and incorporating said weighting term into said deformable part model and employing a post-processing step utilizing a detected landmark point.

19. A non-transitory processor-readable medium storing computer representing instructions, when executed by a processor cause a process of side window detection, said computer code comprising code to:

generate a deformable part model with respect to an image of a vehicle captured in a near-infrared band utilizing a side window detection and a B-frame detection module in order to obtain a set of candidate side-windows;

generate a refined deformable part model utilizing superpixel generation and a longest-line detection in order to remove a false alarm with respect to said deformable part model; and

refine detection performance of said refined deformable part model based on global regional information utilizing a local self-similarity based metric.

20. The non-transitory processor-readable medium of claim 19 wherein said code further comprises code to identify a spatial feature associated with a presence of an occupant.