Return to search

Principal Component Analysis on Fingertips for Gesture Recognition

To have a voice link with other diving partners or surface personnel, divers need to put on a communication mask. The second stage regulator or mouthpiece is equipped with a circuit to pick up the voice of the diver. Then the voice is frequency-modulates into ultrasonic signal to be transmitted into water. A receiver on the other side picks up the ultrasonic signal and demodulates it back to voice, and plays back in diver's earphone set. This technology is mature but not widely adopted for its price. Most divers still use their favorite way to communicate with each other, i.e. DSL (divers' sign language.)
As more and more intelligent machines or robots are built to help divers for their underwater task, divers not only need to
exchange messages with their human partners but also machines. However, it seems that there are not many input devices available other than push buttons or joysticks. We know that divers¡¦hands are always busy with holding tools or gauges. Additional input devices will further complicate their movement, also distract their attention for safety measures. With this consideration, this paper intends to develop an algorithm to read the DSL as input commands for computer-aided diving system.
To simplify the image processing part of the problem, we attach an LED at the tip of each finger. The gesture or the hand sign is then captured by a CCD camera. After thresholding, there will only five or less than five bright spots left in the image. The remaining part of the task is to design a classifier that can identify if the unknown sign is one from the pool. Furthermore, a constraint imposed is that the algorithm should work without knowing all of the signs in advance. This is an analogy to that human can recognize a face is someone known seen before or a stranger. We modify the concept of eigenfaces developed by Turk and Pentland into eigenhands. The idea is to choose geometrical properties of the bright spots (finger tips), like distance from fingertips to the centroid or the total area of the polygon with fingertips as its vertices as the features of the corresponding hand sign. All these features are quantitative, so we can put several features together to construct a vector to represent a specific hand sign. These vectors are treated as the raw data of the hand signs, and an essential subset or
subspace can be spanned by the eigen vectors of the first few large corresponding values. It is less than the total number of hand signed involved. The projection of the raw vector along these eigen vectors are called the principal components of the hand sign. Principal components are abstract but they can serve as keys to match the candidate from a larger pool. With these types of simple geometrical features, the success rate of cross identification among 30 different subjects' 16 gestures varies to 91.04% .

Identiferoai:union.ndltd.org:NSYSU/oai:NSYSU:etd-0731103-094122
Date31 July 2003
CreatorsHsu, Hung-Chang
ContributorsChau-Chang Wang, Chi-Cheng Cheng, Hsin-Hung Chen
PublisherNSYSU
Source SetsNSYSU Electronic Thesis and Dissertation Archive
LanguageCholon
Detected LanguageEnglish
Typetext
Formatapplication/pdf
Sourcehttp://etd.lib.nsysu.edu.tw/ETD-db/ETD-search/view_etd?URN=etd-0731103-094122
Rightsunrestricted, Copyright information available at source archive

Page generated in 0.002 seconds