| Summary: | In several conditions, teleoperated robot system is very useful in solving
severe task in which autonomous robot cannot. Commanding in robot
teleoperation system can be done in several ways, including the use of hand
gesture. Hand gestures are often used as natural interface between human and
robot. Hand shape can gives information in recognizing hand gesture.
Recognizing hand gesture involves several steps.
In this research, the use of Centroid distance Fourier Descriptors (CeFD)
as hand shape descriptor in hand gesture recognition from visually captured hand
gesture is presented. The hand gesture recognition will be used as robot
teleoperation command. Five static gestures, which are adopted from ASL
fingerspelling, are employed in this research. A gesture dictionary is built as a
reference of each gesture. A combination of Nearest Neighbor (NN) classifier and
several distance metrics as gesture recognition system are investigated.
Hand gesture acquisition utilizes Microsoft Kinect as a depth imager. An
analysis of human posture dimension is performed to specify several parameters,
which are used to threshold the acquired depth image. By finding the centroid of
human location in the image, the position of left and right hand can be located.
Thus, both hand images can be segmented. Centroid distance signature is
constructed from the segmented hand contours as a hand shape signature. Fourier
transformation of the centroid distance signature results in fourier descriptors of
the hand shape. The fourier descriptors of hand gesture are then compared with
the gesture dictionary to perform gesture recognition.
Several experiments show the presented system performances. In the
segmentation step, the thresholding operation results in completely segmented
hand images. This segmentation method has low computation time and works
well with an assumption that human hands are positioned ahead of human body.
The use of 15 fourier descriptors as feature vector and Manhattan distance in
Nearest Neighbor classifier achieves the best recognition rates of 95% with small
computation latency of 6.23556 ms. Real-time implementation of the presented
system is applicable due to its processing speed is faster than the imager
acquisition rate.
|
|---|