가속도 센서를 이용한 동적 손 제스쳐 인식

Abstract

Keypad based input method has been widely used for mobile device, such as mobile phone, MP3 player and remote control of electronic devices. As devices has more functions, keypad become more complex. Demand for intuitive interface and recent improvement of sensor and embbedded system technology enable gesture-based interface for mobile device. There are some research interests in gesture recognition with mobile device that contains accelerometer and gyroscope.More atention has also been paid to gesture recognition with an accelerometer only, due to the relatively bulky size and high price of a gyroscope.Isolated gesture recognition is a kind of gesture recognition problem, which is to recognize single gesture from given data. It reqires a kind of pre-processing step to identify block of data involving single gesture before recognition process, whereas continuous gesture recognition system does not. This thesis hasthree main contribution on accelerometer based gesture recognition. First, we suggested an accelerometer based isolated gesture recognition system. Gravity acceleration components are removed form 3-d acceleration signals. Then,blocks of acceleration signals, involving single gesture, are extracted. Extracted blocks are applied to gesture model, to calculate likelihoods. Likelihoods for each gesture models are adjusted by corresponding weights. Weighted valuesof likelihood are finally used to determine a gesture model that best matchesthe input data in the recognition system. Hidden Markov models (HMMs) areused to model each gesture. In contrast other HMM-based approach, we usecontinuous LR HMMs and modified recognition criterion istead of discrete ergodicHMMs and maximu likelihood criterion.Second, we proposed an accelerometer based continuous gesture recognitionsystem. Our HMMs-based isolated gesture recognition system can beexteneded continuous gesture recognition system by merging each HMMs toone huge HMM. This approach can be seen as 2-layered HMMs, the 1st-layermodeled the transition between gesture models, and 2nd-layer modeled eachgesture. Maximum likelihood criterion, used in isolated gesture recognitionsystem, can not be used to predict newly incoming acceleration signals in continuousgesture recognition system. Instead of the criterion, ？？-based approachis used to predict test data.Finally, we suggest pseudo velocity signals, which is transformed from acceleratioinsignals, for improvement of discriminative power. If we transform accelerationsignals to other signals which represent velocity, including directionand speed of motion, it will be helpful to recognize the gesture. Accumulationof acceleration signals can be used for approximation of velocity signals. Butin this case, noise of acceleration signals is also accumulated. As noise levelof acceleration signal is high, due to lack of rotation information and imperfectgravity component removal, noise level of estimated velocity is extrimly high.We adapt template based approach, which is free from noise accumulation.