The Research Of Gesture Recognition Technology Based On Kinect

Gesture recognition is an important as well as a challenging research topic in the field of computer,and as an important part of human-computer interaction,its development affects the naturalness and flexibility of human-computer interaction.However,in the feature extraction,traditional gesture recognition schemes that take color or depth images as research objects are often disturbed by skin color,complex backgrounds,and overlapping occlusions et al.These factors make it difficult to accurately separate gesture images.In the classification stage,traditional classification networks have disadvantages such as large sample demand and low recognition rate.Therefore,in this thesis,the bone information obtained by Kinect is used as a research object to reduce the impact of application scenarios on feature extraction;Feature extraction schemes are designed for the two gesture and motion recognition problems,namely: artificial design schemes to extract geometric features and dual-channel Convolutional Neural Networks to extract random features;In addition,a Long Short Term Memory Network is used to construct a classification model to recognize gestures.The main contents of this thesis are as follows:(1)Gesture recognition can be regarded as the analysis of time series.This thesis uses LSTM to model the context information of the sequence.At the same time,a multi-level LSTM stack classification model was constructed in order to fuse multiple time-scale information of gesture sequences on a global scale and realize high-level abstraction of input data.Finally,by comparing the experimental results of stacks of different levels,it is determined that the four-level LSTM is the best stack level.(2)For the recognition of gestures formed by hands and arms,this thesis uses skeletal information to characterize gestures and designs three geometric features to describe gestures.The experimental results show the effectiveness of the geometric features and recognition framework designed in this thesis in recognizing such gestures.(3)In order to solve the problem of complex gesture recognition including hand shape,this thesis takes the hand-region image and the color-coded image of the spatiotemporal features of bone information as the input of dual-channel CNN to extract high-level random features,and builds a multilayer LSTM gesture based on random features Identify the framework.At the same time,in order to solve the problems of small and confusing hand regions,a Multi-scale Faster Region-based Convolutional Neural Network is proposed to obtain hand region images.The experimental results show that the above method can well handle the hand region acquisition and feature extraction in complex gesture and motion recognition,and it achieves a recognition rate of 97.834% on a custom data set.The results of evaluation experiments of various recognition methods on large public "NTU RGB + D" data sets are compared and analyzed,and the priority of the designed scheme is verified.