1 results listed
In this paper, one-shot-learning gesture recognition
methods are reviewed and an approach of hand gesture
recognition using one-shot-learning is proposed. This approach
aims to recognize new categories of gestures from a single video
clip of each gesture. The gestures are generally related to a
particular task, for instance, hand signals used by divers, finger
codes to represent numerals, etc. In this study, both RGB and
depth images are utilized for a given dataset. A rich dataset,
namely the ChaLearn Gesture Dataset (CGD2011), are employed.
The dataset is divided into 20 different files which include 940
videos in total. Although training the system with only one
example is difficult, depth and RGB images provide many new
possibilities. We used the standard deviation of the depth images
of a gesture and motion history image (MHI) method. Also, two
dimensional fast fourier transform (2D FFT) is used to reduce the
effect of camera shift. It is seen that FFT has no distinct effect on
the image quality. Then, we compare image templates based on the
correlation coefficients and Levenshtein, Mahalanobis, Frobenius
distance measures. The Levenshtein distance measure is more
suitable to match image templates compared to other distance
measures. It is observed that MHI method gives better hand
gesture recognition accuracy about one-shot-learning.
International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES
Esma Şeker
Oğuz Findik