Human Action Recognition Method Based On Neural Network And Skeletal Points Features

In recent years,action recognition has been widely concerned by researchers,and has been widely used in security,medical,and sports event analysis.At present,action recognition based on RGB video has been extensively studied,but the acquisition of RGB images is susceptible to illumination,angle,etc.,and RGB image information is difficult to represent spatial changes in motion.With the development of depth sensors and the research of human pose estimation algorithms,the acquisition of human points coordinates becomes easier.The joints coordinates represent the 3D position information of human body parts in space,which is more conducive to motion recognition,so it is based on bone points.Action recognition has become a hot topic in the field of motion recognition.The process of Skeleton-based action recognition first extracts the spatiotemporal features in the action sequence,then sets the classifier to classify the extracted features to obtain the action recognition result.The pose of each frame in the action sequence has spatial correlation,and the different inter-frame poses have temporal correlation.It is difficult to extract spatiotemporal features simultaneously with common geometric features or time series features.In order to solve this problem,We propose a method for extracting spatial features based on skeleton joint.In order to enhance the expression of space-time features by classifiers,we propose a novel deep stacked Bidirectional Long Short-Term Memory(DSB-LSTM)network for skeleton-based action recognition.The main research contents of this paper are as follows:1)A method based on skeletal joint feature extraction is proposed.Firstly,a set of skeletal vectors is constructed according to the human skeleton geometry.Then,the skeletal vector modulus ratio and the skeletal vector angle feature are extracted by this set of skeletal vectors to verify the structural features of the skeletal points.Effectiveness,this paper sets up the LSTM neural network as the model classifier,conducts experiments on the public dataset,and compares the extracted skeleton pointfeatures by setting a variety of classifiers.2)The action sequence has time series,and the above-mentioned single LSTM neural network as the classifier can not construct a good time-series expression of the extracted skeleton joints.In order to enhance the feature recognition effect of extraction,this paper proposes a DSB-LSTM.The internet.Based on the structural features of the skeleton points,the method is based on stacking multiple deep bidirectional LSTMs,setting the robustness of the Masking layer enhancement model to the missing data,setting the timing dropout layer to reduce the model overfitting,and setting the maximum timing pool.The layer enhances the expression of time series features.3)The above proposed skeleton based spatial feature extraction method and DSB-LSTM networ are experimentally tested,and tested on MSR-Action3 D,Forence3D-Action and UTKinect-Action three public data sets,then Comparing the schemes,the action recognition accuracy on the test data set has different degrees of improvement.The experimental results show that compared with the existing algorithms,this paper proposes that the feature extraction method and the bidirectional LSTM neural network model can effectively improve the accuracy of motion recognition based on bone points.