Gesture Recognition System Based On Complex Background

With the development of the technology,the human-computer interaction techniques become more and more important.Gesture recognition is the main form of the human-computer interaction techniques.It's principle is using the computer to process the image information which containing the gesture to identify the external shape of the gesture.As a special part of the human-computer interaction techniques,it can not only enhance the human-computer interaction capacity,but also improve people's interactive experience.This design has made a comprehensive interpretation and analysis of the gesture and its related recognition technology in recent years,and carefully studied the technical difficulties of the gesture recognition.When completing the design and development of the gesture recognition in PC and other embedded platform,the most important part is analyzing and studying deeply to the feature extraction and recognition of gesture recognition technology.This design mainly contains following parts:First of all,we designed a whole structure based on the requirements of the practical application.This system contains two parts: static gesture recognition and dynamic gesture recognition.A large number of samples were set up for training and testing.As for the most important parts: the tracking segmentation,feature extraction and gesture recognition of the gesture,we mainly designed following parts: Firstly: We combines the movement information and the classic Camshift algorithm to track the gestures.Secondly: We preprocessed the gesture region image which been captured,and used the skin color to detect the segmentation of gestures.Thirdly: After extracting the gestures' Histogram of oriented gradients characteristics,we respectively using KNN,SVM,KNN_SVM,RF to recognize,and we found that combining the KNN_SVM is the highest recognition rate.Fourthly,after analyzing and verifying the identification results one by one,we found that the HOG feature can't recognize the similar hand type.After trying other new features,we found that using KNN_SVM to recognize the HOG_LBP features have an obvious improvement in recognizing the similar hand type.The average recognition rate was 97.2125%.The method is also validated on Marcel library and has a better effect.Finally,we combined the Kalman and Camshift tracking method to finish the dynamic gesture recognition tracking.And then extract the gestures centroid,which can get the gesture trajectory.We used 12 degrees coding to discretize orbit characteristics according to the gesture trajectory.We send the extracted gesture trajectory into the SVM for gesture recognition,the average recognition rate achieved better effect.After transplanted into and tested in hardware platform Raspberry Pi 3b,we found this system can realize real-time gesture recognition,which complishes the design target.moreover,this system is stable and the comprehensive recognition rate can achieve 94.12%.The real time recognition meets the requirement of usage,the development software can realize the automatic boot and automatic boot the camera and real-time recognition.This system has certain application value.