Research On Static Hand Gesture Recognition Based On Convolutional Neural Network

With the extremely rapid progression of automation and intellectualization technologies nowadays,human society has been constantly demanding better and better Human-Computer Interaction(HCI)interfaces implicitly.As a communication approach which have been used way back in the beginning of human history,hand gesture has the characteristic of being both intuitive and effective,and as a consequence,has become an indispensable component in modern HCI devices.It would be obvious that a stable,efficient and effective hand gesture recognition system is the key to any communication interfaces based on hand gesture control,which almost entirely determines the usability and user experience of said interface.The traditional computer vision method of gesture recognition often involves two parts: feature extraction and classification,in which the feature extraction part usually has an influence directly on the final performance.The effectiveness of hand-crafted features is vastly related to the experience of domain experts,and because of that,uncertainty is almost inevitable.Also,the feature engineering process is both costly and error-prone.The fact that a great amount of work needed to be done in order to increase a tiny bit of performance is seriously limiting.Recently,due to its superior performance compared to other conventional methods in various computer vision research areas,deep convolutional neural network's capability of feature extraction and generalization have been recognized gradually.Since in essence,convolutional neural network combines the process of feature extraction and classification into an end to end procedure and can learn those parameters by itself,which requires very little human interaction,it is particularly suitable to construct different recognition systems for different gestures sets.In this paper,we apply the theory of convolutional neural network to the field of static hand gesture recognition,in which the main works can be summarized below:(1)Inspired by all-convolutional neural network,two concise and effective convolutional neural network models have been proposed.By comparing the accuracy with some of the related work on several hand gesture datasets,we show the advantages of convolutional neural network in such tasks.(2)With the motivation of further improving the performance and effectiveness of our models,we compared and analyzed two methods of ensemble learning to combine our base models.The resulting model has better accuracy compared to any of the single base neural network model in use,which,thanks to the reduced computation complexity of base models,can still run in real time for the requirement of a real world recognition system.(3)To ease the painfully process of collecting gestures from a diversity of signers or volunteers systematically,a web-based dataset collection and management system has been developed with the goal of reducing the necessary communication between these two ends.And by using it,a hand gesture dataset to test our models extensively has been made.It consists of 14 kinds of gestures and more than 8000 samples with driving car control functionality in mind,hence the name MCSHG(Motion Control Static Hand Gesture).