Gesture Recognition Ideas and Exploration

2stage Recognition of Raw Acceleration Signals
Relys on on or off states
Recognition Framework (
1 . Calibration,
2. Preprocessing
1. Motion area detection: interval of a gesture motion in time
2. Norm Normalization: Allowing for gravity
3. Gaussian Smoothing: Remove hand trembling
4. Resampling: Normalise in respect to motion speed
3. Feature point Extraction: determine the distance between the sampling points
4. Recognition:
1. Bayesian Network modeling of relationships
2. find model with highest model likelehood
5. Confusing pair discrimination
1. Determine between confusing models I.E. 0 and 6
2. Double intergration doesn't work becaue of double intergration error accumulation

Chapter 10. Gesture Recognition, Mathew Turk
Hummels and Stappers 4 aspects of a gesture wich may be important to its meaning
1. Spatial: where it occurs, location gesture refer to
2. Pathic: the path which a gesture takes
3. Symbolic: the sign a gesture makes
4. Affective: the emoitional quality of a gesture
Sturman suggested a taxonomy of whole-hand input that categorizes input techniques along 2 dimnentsions
1. Classes of hand actions: continous or discrete
2. Interpretation of hand actions: direct, mapped, or symbolic.
System Design Suggestioncs
1. Do inform the user
2. Do give the user feedback
3. Do take advantages of uniqueness of gesture
4. Do understand the benifits and limitations of the particular technology
5. Do usability testing of the system
6. Do avoid temporal segmentation if fesible
7. Don't tire the user
8. Don't use gesture as a gimmick
9. Don't incrase the user's cognitive load
10. Don't require precise motion
11. Don't create new, unnatural gestural language

Hidden Markov Models for Spatio-temporal pattern recognitions
Conclusions
1. Baum-Welch algorithim is a hillclimbing technique that is generally unable to fiund the gobal maxima
2. Baum-Welch is often very good
3. HMM perform better than Baum-Welch on simulated HMM data, but htese result do not nessecarily translate into improved perfromance in real world applications.
Other Points
1. Fully connected topology: There is not necessarily a defined starting stte and all state transitions as possible such the Aij |= 0 V i,j E[1,N]
2. Left-Right Topology: popular in speech recognition application. There is a defined stateing stae and only state transitions to higher-index state are allowed
3. Left-Right-banded topologies: transition structure contains on;y self-transitions and nnext state transitions, I.E. state cannot be skipped
3. Gestures have a natural state and finish state and thus it is reasonable to adopt the LR model


When to recognise
Many applications require an on off state to be selected before gesture recognition becomes avaliable. Stoke-it relys on the right mouse being pressed down. Hollar et al Wireless statis hand gesture recognition suggests movement from the hand from stble and the iPhoine activates when the screen is touched.


From AiLive movie
Problems come from speed, orientation, and direction, left/right handed. Ussually uses 3 motions