SIBGRAPI 2005


People detection in still images based on a skin filter and body part evidence

Claudio Cavalcanti, Herman Gomes

Universidade Federal de Campina Grande
Departamento de Sistemas e Computação

Abstract. This paper presents an approach for determining the presence of people in a still image. The approach starts with the segmentation of skin areas in the images using the YCbCr color space and a set of experimentally defined thresholds. The detected skin areas are then grouped into body part candidate regions. Regions that are too small (typically smaller than 5x5) are discarded. A simple heuristic rule, based on the proportions and pixel density of these regions, is then applied to map these regions into body part evidence. The application of the proposed approach to a database of still photographs resulted in the correct classification of 77.2% of the images.

Parternship. This project was developed with the collaboration of Hewlett-Packard.

[ contact authors ]


1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18


Figure A.1: Sample of images without people.




1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18


Figure A.2: Sample of images with people.




1

2

3

4

5

6

7

8

9

10

11

12

13

14

15


Figure A.3 Exemples of correct detection of people (see Fig. A.2 for reference, the last 3 images in Fig. A.2 were incorrectly classified as not having people).




5

1

16



Figure A.4 Examples of false detection (see Fig. A.1 for reference).