Enhanced Gesture Recognition Performance through Improved Pre-Processing

Abstract

Gesture recognition is analyzed on a set of static hand gestures in the context of designing robust, real-time pre-processing techniques for applications in hand-held electronics. A comparative case study that uses various combinations of algorithms across the steps of the recognition process is made, revealing the fact that many method combinations can produce highly accurate results, even at low resolutions, given the right kind of pre-processing. The pre-processing includes the hand segmentation and normalization done before feature extraction. Indeed, pre-processing has by far the greatest effect on the overall accuracy, robustness, and speed of the gesture recognition process, significantly outweighing the influence of feature-extraction and classification. Even at image resolutions as low as 8×8 pixels, accuracies of 99% are achieved using a simple PCA feature selection scheme and a LDA classification method. These results suggest the priority and advantages of focusing on developing robust and efficient pre-processing methods.

Selecting a Small Set of Optimal Gestures from an Extensive Lexicon

“…The hands have already been isolated within each image and segmented with black backgrounds. The raw images are pre-processed by converting to grayscale, centering the hand within the frame, down-sampling the images to 32 × 32 (without changing the aspect ratio of the hand), and normalizing the brightest pixel intensity [28]. As has been stated, feature selection is done by the PCA method and the GP method, both only using 10 features.…”

“…As has been stated, feature selection is done by the PCA method and the GP method, both only using 10 features. Statistical learning will be done by the LDA method [29], with a one-vs-the-rest classification style [28]. The training set consists of 20 randomly chosen images from each gesture class, and the test data consists of the remaining 50 images (45 for class t) from each class.…”