Gesture recognition using auto-regressive coefficients of higher-order local auto-correlation features

We propose an efficient method for motion recognition from time-varying images. The method extracts higher-order local auto-correlation (HLAC) features from time differential images in a time series. An auto-regressive model is applied to each dimension of the HLAC features and AR coefficients are calculated to achieve efficient extraction of information over a time range within a (time) window. These features become high dimensional feature vectors, so we use discriminant analysis to obtain effective less-dimensional features. These features are learned by the hidden Markov model (HMM) based recognizer to cope with non-uniformity of the speed of motions. We applied this method to gesture recognition and obtained good results. Because HLAC features are location-invariant, our method is robust to position changes of the objects (humans) in the image and need no segmentation of the objects. In addition, since the method extracts features from time differential images, it is also robust to changes of background and illumination condition. The method does not necessitate any a priori knowledge about objects in images; therefore, it may be applied to various applications of motion recognition, such as lip-reading, sign language recognition, and so forth.