Human Action Recognition Based On Depth Sequence

Human motion recognition is important in many applications,such as smart security,machine monitoring,human-computer interaction,virtual reality,driverless,and more.With the continuous development of artificial intelligence and pattern recognition technology,the huge industrial prospect of human motion recognition has made it a research hotspot in recent years.The ensuing hardware device innovation provides more possibilities for the field of human motion recognition.The planar information is upgraded to spatial information.More and more research is beginning to invest in the direction based on the depth image sequence,while in the depth image In the sequence,3D skeleton data is most widely used.At the same time,with the continuous development of deep learning technology,the traditional manual feature extraction method has gradually been replaced.The motion recognition method combining deep learning and 3D skeleton data has achieved good performance in experimental results.Some researchers' practice is to use a convolutional neural network to extract features from skeleton data in individual frames to determine the action category.The limitation of this method is that it only pays attention to the skeletal posture at a single moment,and seriously ignores the influence of the features on the time dimension on the motion recognition performance.At the same time,the feature extraction of the spatial structure of the skeleton itself is limited,and the joint data is treated as independent.The relevant data is trained,and the influence of the topological relationship of the human body structure on the recognition result is lost,and it is highly susceptible to interference from the noise joint.In view of the above problems,this paper aims to propose a network model that can extract high-quality time and space features for 3D skeleton data to improve the accuracy of motion recognition.The main research work of this paper is as follows:1.In the aspect of spatial feature extraction,based on the characteristics of human motion system,after reasonable calibration and normalization of 3D skeleton data,the method of using the motion vector of joint and parent joint as training input is proposed.Regularity can avoid the effects of noise joints.At the same time,it is proposed that the human skeleton joints are divided into the upper body joint and the lower body joint according to the topological relationship,so that different actions can obtain different weights in the upper body and the lower body.Compared with the motion characteristics of the whole human skeleton,subdivided into the motion characteristics of the joint and the parent joint and the motion characteristics of the upper body and the lower body,the robustness and recognition accuracy of the algorithm are improved.2.In terms of time feature extraction,this paper proposes a multi-level long-term and short-term memory(LSTM)network structure for motion recognition.LSTM is better able to extract features in the time dimension than other deep neural networks.The single LSTM network structure has limited effect on skeleton space feature extraction.The multi-level LSMT network proposed in this paper is based on 3D skeleton data characteristics and human body structure features,from joint data processing in fine-grained sub-networks to upper-body joint characteristics in medium-grained subnets.Furthermore,the relevant features are further extracted,and finally the whole structure is fused with the entire skeleton feature,and the motion recognition is performed from part to whole to obtain higher recognition accuracy.In this paper,the two methods proposed above are used to perform cross-body experiments on the common data set NTU RGB + D.The test results verify the performance effectiveness of the proposed method.