Real-Time Tracking and Simulation of Multi-Modal Human Action under Deep Learning Algorithm
The diversity of human motion poses and the impact of factors such as light,shadow and noise on the extraction and analysis of human motion characteristics are significant.If there is an obstruction in the tracking process,it is easy to cause target loss,seriously affecting the accuracy of human motion tracking.In order to address these issues,an autoencoder in the deep learning method was introduced,and a real-time tracking method for multi-modal human motion was proposed.By processing human motion images,we extracted rough features of human motion.Then,we used the autoencoder to enhance these features,thus achieving the fine extraction of human motion features.Moreover,we constructed a template including conventional human motion features.Furthermore,we used kernel correlation filters(KCF)to obtain the response function peak and establish a target loss criterion.In combina-tion with the feature matching algorithm,we calculated the similarity between the target and regional structural fea-tures,thus determining the best tracking target.Finally,we achieved the real-time human motion tracking.Experi-mental results show that the proposed method has high precision and short latency,and the tracking effect is good.
Deep learning algorithmMultimodal human motionReal-time tracking methodHuman motion char-acteristicTracking algorithm
NETL
NSTL
万方数据
