The Research And Application Of Hand Gestures Recognition Based On Machine Learning

In recent years,the development trend of smart wearable devices has been very rapid,such as smart watches,which have become a common device in people's daily life.At the same time,human behavior recognition based on wearable device sensors has become a new research direction in the computer industry,due to its rich content perception,strong computing power,real-time acquisition of human motion state information and the advantages of less privacy invasion compared with image-based behavior recognition.This thesis aims to conduct the research of hand gestures recognition and smart home control based on machine learning and smart wearable devices.The gestures are classified by analyzing and processing the sensor signals generated by human motion and combining the classification algorithm models designed in advance.Based on the research and analysis of users' daily gestures,this thesis constructs a gestures dataset,and analyses and processes it.The improved algorithm for curve feature extraction has been exploited to extract the complete gestures sensor data,and then the data format is processed by the template in this thesis.Finally,the proposed convolutional neural network has been trained with the data set of hand gestures to perform the recognition of hand gestures.12 types of hand gestures are collected by the popular commercial smart watch in this thesis.And this thesis attempts to classify 12 kinds of hand gestures and complete the perception of hand gestures such as writing Chinese characters and Arabic numerals.The main research in this thesis is divided into two parts,which are the preprocessing of sensor signals for hand gestures and the classification of hand gestures.In the first part,the acceleration and gyroscope signals of hand gestures are collected and processed,and the improved curve feature extraction algorithm is proposed to extract the complete hand gestures sensor data.At the same time,a set of data format template processing scheme is designed to format the input data of the classification model.In the second part,a new convolutional neural network is proposed to classify the hand gestures based on the actual application scenarios and the computing ability of the equipment.The application of multi-time series and partial weight sharing technology effectively improves the recognition accuracy of the algorithm in this part.And the backpropagation algorithm of enabling the convolutional neural network to perform correct training in this thesis.In this thesis,1800 sets of action data are constructed,in order to verified and compared the proposed algorithm.The results show that the proposed algorithm achieved good recognition accuracy,which is superior to the traditional recognition method(decision tree)and the traditional convolution neural network.Finally,this method has been applied to practical application,and the code of implementation for this method is given.