The Robot's Human Action Recognition And Imitation Based On Machine Learning

Humanoid robot can communicate with people just like human beings,for they also has "vision","hearing","touch" and "emotion",and it's meaningful that make a humanoid robot have a interact communication with people or other robots efficiently and naturally.The interaction between humans and meachines that based on visual is more in line with the human habits,which also make people feel good.The research for training process simpler and more accurate recognition effect is current hotspot,and machine learning methods are of great help to this problems.Therefore,this paper is focus on human action recognition and imitation with machine learning methods.This paper summarizes the machine learning classification and the typical algorithm,feature extraction and human action recognition methods research,and proposes a fused feature extraction based on these methods research.In addition,this paper proposes a human action recognition method which combines dynamic recognition with static recognition,finally,achieve the human action recognition and imitation method in NAO through the coordinates and the joints'radian.The main work includes:(1)The target detection which use the background difference method based on probability and statistics for the static background and the slowly changing background,provides a foreground target,then lock on this foreground target by camshift algorithm,output the outline imformation of foreground target and the foreground changement of target's coordinate information.(2)Accoeding to Hu moment's invariance property of rotation,translation and scaling changement,put forward a fused feature extraction combines the overall relative geometric features with Hu moment.Based on the information of foreground target's outline,chose eight overall relative geometric features,which includes the length-width ratio,rectangle ratio,area and convex surface ratio,perimeter and convex perimeter ratio,the globular features ratio,internal and external circular ratio,eccentricity ratio and the shape parameters,combined with Hu moment output a fused feature,that can improve the human action recognition efficiency.(3)Using machine learning discriminant model of SVM classifier as the classification model of action recognition,training the discriminant function of multi class SVM classifier through gaussian kernel,and recognizing the the static posture based on the fused feature,at last,analysis of the human action with the foreground target's coordinate information,namely the human action recognition method which combines dynamic recognition with static recognition.(4)Put forward a radian's computing method based on the overall relative geometric features.Import the human action recognition results,the coordinates and the joints'radian into NAO,achieve the human action imitation.