A JSL sentence consists of sign words represented by hand motion. Each word is articulated into hand motions of the both hands, then each one-handed motion into changes of the hand configuration consisting of hand shape, orientation and position. Sign words can be recognized, if both one-handed motions are understood. In the following, we focus on recognition of one- handed motions.
Every hand motion is repetition of motion and rest. According to our observations, in motion periods the hand trajectory is much more important in recognizing sign gestures than the hand shape, and in resting periods the hand configuration at the beginning and the end of the period are essential. Therefore, every one-handed motion should be described by both a sequence of the hand trajectories in motion periods and a sequence of each element of the hand configuration in resting periods.
Because sign words are combined continuously into a JSL sentence, detection of word borders is required for word recognition. Although this segmentation problem is challenging for fluent JSL, it is not discussed more in this paper, because we are dealing with relatively slow and "legible" signs that can be split at a long pause. A preparatory motion appears before each word starts, which is the transition from the final hand state of the preceding word to the starting state of the next. The preparatory motion varies according to the both words. We include this motion in the following word for convenience.
A VPL DataGlove Model II is used for acquiring hand data. It has two sensors for measuring bending angles of the two joints on each finger, one over the knuckle and the other over middle joint of the finger. There is also a "Phloem's" sensor attached to the back of the glove which measures three position data and three orientation data (roll, pitch, yaw) relative to the fixed magnetic source. The position data is calibrated by subtracting the neutral position data from the raw position data; they are normalized with the breadth of signer's shoulders. 16 pieces of data per hand are measured in the coordinate system drawn in Fig. 1 at each sampling instance (30 times/sec).