CHOI, Yoo-Joo¹, KIM, Ku-Jin², CHO, We-Duke³
1 Dept.of Computer Science and Application,Seoul University of Venture and Information, Seoul, Korea, email@example.com
2 Dept.of Computer Engineering, Kyungpook National University, Daegu, Korea, firstname.lastname@example.org
3 Center of Excellence for Ubiquitous System, Ajou University,Suwon, Korea, email@example.com
This paper presents a grid-based approach for robustly extracting head and hand regions of a moving human in a varying distance from the camera. First, our method applies background subtraction scheme to the sequence of images based on hue and saturation information to classify the foreground and background pixels. While the original given images are with 648468 resolutions, the background subtracted image is partitioned into grid patches, where each grid consists of 8 8pixels. The grid patches are classified into background, non-skin foreground and skin foreground classes based on the histogram analysis of patch feature values. The histogram analysis of patch feature values makes patch classification be robust regardless of the distance from the camera. Then, the connected component labeling is applied to the grid image which consists of the classified patches. By using the grid image, we can effectively extract the skin regions of human head and hands, and we also can reduce unexpected labeling results from the noises in detecting the skin regions.
Keywords: human-computer interaction, gesture interface,motion detection, background subtraction
Background subtraction results
Fig.1. Background subtraction images in 648468 resolution. Non-Skin foreground pixels are colored red and skin foreground pixels are colored yellow.
(a) Subject in a long distance (b) Subject in a short distance
Fig. 2. The ratio of the numbers of foreground and skin pixels in the images in different distances. (a) The ratio of foreground pixels to total pixels is 10.73 % and the ratio of skin pixels to foreground pixels is 3.5 %. (b) The ratio of foreground pixels to total pixels is 20.2 % and the ratio of skin pixels to foreground pixels is 11.2 %.
(a) Subject in a long distance
(b) Subject in a short distance
Fig. 3. Histogram of the ratio of foreground pixels in a grid (Hfg) and histogram of the ratio of skin pixels to foreground pixels in a grid(Hsk).
Fig. 4. Extracted grid images in 8154 resolution. General foreground patches are colored blue and skin patches are colored pink.
Using static threshold
Using adaptive threshold
Fig. 5. Comparison of skin ROI extraction results of between static patch classification and adaptive patch classification. Yellow rectangles depict the skin regions of interest.
Fig. 6. Skin ROI extraction result for a person wearing a short sleeve shirt. (left) original image, (middle) detected skin regions, (right) magnified image.
Maintained by firstname.lastname@example.org