LSTM-based Human Continuous Motion Recognition

With the development of the machinery manufacturing industry,intelligent robots using human motion recognition technology provide practical possibilities for efficient human-machine collaborative assembly.By recognizing and understanding human actions,robots can work better with humans,thereby greatly improving production efficiency.In addition,human motion recognition is also widely used in many fields such as smart security,smart home,smart medical and other production and life fields.In this paper,three main aspects of action data feature extraction,action recognition model construction,and segmentation recognition of human continuous motion are implemented to realize human motion recognition research.The human movement model was established.Mean filter processing is performed on the original 3D skeleton data of the human body to eliminate possible data noise.Based on the simplified model of the human skeleton composed of the main joint points,the static features composed of the angle of the limb and the relative distance,the dynamic features composed of the joint kinetic energy and the angular acceleration of the angle of the limb are established,and the static and dynamic features are carried out Feature fusion to characterize human movements.Through the key frame extraction model,the action frames with obvious changes in the action sequence are selected,and the action frames with insignificant changes are eliminated,so as to improve the accuracy of action recognition and reduce the amount of calculation.Two human motion recognition models based on LSTM neural network and Bi-LSTM neural network with attention mechanism and Dropout are constructed.First,LSTM neural network is used to classify human actions,and then Bi-LSTM neural network is used to introduce attention mechanism and Dropout to make up for the shortcomings of LSTM neural network in human action recognition.Bi-LSTM neural network can be extracted to be more comprehensive Information,attention mechanism to increase the focus on the main information,Dropout to prevent overfitting of the neural network.Then compare these two motion recognition models.The orthogonal test method was used to optimize several main parameters in the neural network,and a high recognition rate was finally achievedThe continuous motion recognition method is studied,and the continuous motion recognition model is constructed.Aiming at natural continuous motion,two human body continuous motion segmentation and recognition methods are presented.An energy-based method for segmentation of continuous human movements is proposed,which uses the general idea that the energy value of human movements is higher than that of transitional movements to complete the division.Then,a method for segmentation and recognition of continuous actions based on sliding window combined with neural network classifier is given The action boundary points in continuous actions are identified and detected,and the final boundary points are demarcated by using the screening mechanism.Compared with other methods,a better segmentation recognition result is obtained.