Robust hand gesture recognition using rapid transform
Date of Issue2015
School of Electrical and Electronic Engineering
In this thesis, a robust and fast hand gesture recognition system is proposed by using the rapid transform (RT) to extract the salient features inherited in various hand gestures. The proposed system is ready to be realized in any mobile device (e.g., smart phones, Google glass TM) that has been equipped with a typical RGB camera. The proposed system has an extremely-low computational and algorithmic complexities (hence, incurring low power consumptions). It also delivers highly accurate hand gesture recognition with real-time performance. So far as the hand gesture pattern variations are concerned, the RT exploited in our system is completely invariant to shift (or translation) and highly invariant to a wide range of rotations, approximately [40, 40] as demonstrated in our simulation results. Although the RT is scale sensitive, this issue can be easily overcome by applying a simple hand-size normalization operation on the detected hand region as part of the preprocessing stage. All these attractive merits greatly facilitate the use of hand gesture recognition in various real-life applications. The processing flow of the proposed system is described as follows. Initially, one template for each gesture needs to be created and pre-stored by the user in advance. This processing flow is the same as that in the hand gesture recognition stage. When a hand gesture is posed by the user and to be recognized by the system in the recognition stage, the hand gesture image performed by the user needs to be reasonably captured by a camera, say, from the user’s mobile device. The hand region will be detected based on skin color. The segmented hand region will be treated as the foreground object, followed by converting it to a binary image version with further hand-size normalization to address the above-mentioned scale issue. Finally, the computed RT is applied to the normalized binary image for extracting salient features of the hand gesture. The RT results are then discriminated against a small set of pre-stored template database based on the nearest-neighbor classifier. Extensive simulation results have clearly shown that our developed RT-based hand gesture recognition system achieves a fairly high recognition rate and is quite robust against practical hand gesture variations, such as translation, scale, orientation, and tilt.
DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems