1. Real-time Hand Pose Recognition Using Low- Resolution Depth Images 2006.10.5

2. outline Introduction Method - A hand model - Segmenting hand, wrist, and background - Locating palm center - Locating obtruding fingers Hand pose recognition Discussion

3. Introduction The difficulties of gesture recognition using regular cameras include the following: - Noisy segmentation of hand from complex backgrounds and changing illumination. -Simple features extracted from images produce ambiguities, whereas sophisticated feature extraction can increase processing time. -Hand tracking methods suffer from initialization and tracking failures.

4. Hand pose representation In this paper, we propose a finger spelling mechanism that decompose a hand pose into finger states, and define the hand pose as finger poses and finger inter-relation.

5. Method

6. Method hand model Hands can be modeled at different level of detail, according to the application needs. A new user is required to initialize the system by producing the pose of a “flat” hand facing the camera.

7. Method segmenting hand, wrist and background A depth value separate the foreground (F) and background. Where is the row index, column index, and depth value of a point in the depth image.

8. Method segmenting hand, wrist and background The boundary that divides the wrist into foreground and background. Where B is the boundary that encloses the F

9. Method Locating palm center The palm center is defined as the point in the hand region that maximizes its distance to the closest hand boundary. Where d2 as the distance in the image plane. The palm size R is defined as the distance between and the closest boundary point.

10. Method Locating obtruding fingers For an obtruding fingertip t, it is apparent that either or Where is a scaling factor, and is a threshold.

11. Method Locating obtruding fingers (profile-based) An obtruding finger region is located using the following operation: -1. compute the curvature of the boundary curve -2. local curvature extreme can be classified as “peak” points and “valley” points; -3. a potential fingertip divides the boundary B into and -4. starting from, for each point on, a corresponding point on is computed as the nearest point to ; and vice versa a point for each on ; -5. terminate step 4 when a) a “valley” point is encountered, or b) distance is a threshold related to the palm size

12. Method Locating obtruding fingers (depth-based) Depth boundaries D are located where a noticeable depth difference exists between neighboring points.

13. Method Classify finger poses and inter-relations Each located finger region is defined by its boundary and average width of the region if is a single- finger region.

14. Hand pose recognition

15. Discussion There are a few hand poses and cases our algorithm fails to recognize