Real-time clothing recognition in surveillance videos
View Patent ↗Systems and methods are disclosed to recognize clothing from videos by detecting and tracking a human; performing face alignment and occlusal detection; and performing age and gender estimation, skin area extraction, and clothing segmentation to a linear support vector machine (SVM) to recognize clothing worn by the human.
1. A method to recognize clothing from videos, comprising
detecting and tracking a human,
evaluating texture features based on histogram of oriented gradient (HOG) in multiple spatial cells, a bag of dense SIFT features, and DCT responses, wherein the gradient orientations on 8 directions (every 45 degree) are computed on color segmentation results;
performing face alignment and occlusal detection; and
performing age and gender estimation, skin area extraction, and clothing segmentation to a linear support vector machine (SVM) to recognize clothing worn by the human.
2. The method of claim 1 , comprising checking for quality of shape/texture and color histograms.
3. The method of claim 1 , comprising evaluating different clothing representations combining color histograms with a histogram of oriented gradient (HOG), Bag-of-Words (BOW) features, and discrete cosine transform (DCT) features.
4. The method of claim 1 , comprising combining gender and age, skin area ratios of limbs, color, shape and texture features using the clothing segmentation results as a mask.
5. The method of claim 1 , comprising using a Voronoi image to conduct seed selection in a region growing segmentation method.
6. The method of claim 1 , comprising determining gradient orientations at every 45 degrees on a color segmentation.
7. The method of claim 1 , comprising spatially dividing top and bottom parts of a human image into cells, and concatenating histograms of all cells as a HOG feature.
8. The method of claim 1 , comprising quantizing local features with a visual codebook and normalizing frequencies of the codewords to generate a BoW descriptor.
9. The method of claim 1 , comprising packing components of DCT coefficients to form a DCT feature.
10. The method of claim 1 , comprising determining local shape descriptors of color segments on the human.
11. A system to recognize clothing from videos, comprising
means for detecting and tracking a human,
means for valuating texture features based on histogram of oriented gradient (HOG) in multiple spatial cells, a bag of dense SIFT features, and DCT responses, wherein the gradient orientations on 8 directions (every 45 degree) are computed on color segmentation results;
means for performing face alignment and occlusal detection; and
means for performing age and gender estimation, skin area extraction, and clothing segmentation to a linear support vector machine (SVM) to recognize clothing worn by the human.
12. The system of claim 11 , comprising means for checking for quality of shape/texture and color histograms.
13. The system of claim 11 , comprising means for evaluating different clothing representations combining color histograms with a histogram of oriented gradient (HOG), Bag-of-Words (BOW) features, and discrete cosine transform (DCT) features.
14. The system of claim 11 , comprising means for combining gender and age, skin area ratios of limbs, color, shape and texture features using the clothing segmentation results as a mask.
15. The system of claim 11 , comprising a Voronoi image to conduct seed selection in a region growing segmentation method.
16. The system of claim 11 , comprising means for determining gradient orientations at every 45 degrees on a color segmentation.
17. The system of claim 11 , comprising means for spatially dividing top and bottom parts of a human image into cells, and concatenating histograms of all cells as a HOG feature.
18. The system of claim 11 , comprising means for quantizing local features with a visual codebook and normalizing frequencies of the codewords to generate a BoW descriptor.
19. The system of claim 11 , comprising means for packing components of DCT coefficients to form a DCT feature.
20. The system of claim 11 , comprising means for determining local shape descriptors of color segments on the human.