| Summary: | 碩士 === 國立成功大學 === 電機工程學系碩博士班 === 95 === This thesis comprises of a programmable human behavior detection scheme based on the principal of state transition support vector machine (STSVM). This model is adopted for the classification of human behaviors and the state classification is done by implementing a programmable human behavior detection algorithm. The proposed module utilizes Viterbi’s model to detect the optimal path in state classification. For validating the proposed system, we defined five events such as (1) raising hand; (2) standing up; (3) squatting down; (4) falling down; and (5) sitting.
This system can be programmed by the users and they can abide the rules to train each event classifier, of some human behaviors. Through this programmable approach easy implementation, flexible modification, real time human behavior detection could be highly achievable. Our approach can be accomplished in the ubiquitous computing environment, which we believe as our highlighting feature. Since every event is time-sliced, 10 images per second are initially obtained from web camera. Through the existence of relative
Probability between the current image and next image a transition probability of these images could be found and state probability of each image shall be obtained by STSVM. These probabilities could be utilized to determine an optimal path, which in turns helps in checking all the paths for event detection. Experimental results prove robustness and effectiveness with a precision rate of 82%, which would be is highly suitable for human behavioral detection.
|
|---|
