Research On Human Motion Recognition Method Based On Deep Learning

Human motion recognition based on machine vision can identify human motion characteristics from a large amount of video data,and relevant research results have provided convenience for people's lives.In order to improve the recognition rate of human motion in video,this paper proposes an improved deep network model to identify and study human motion.The model uses dense optical flow function to process data,using improved two-dimensional convolutional neural network(2DCNN),three-dimensional Three network cascade models of convolutional neural network(3DCNN)and long-term and short-term memory neural network(LSTM)are used to extract human motion feature information.After the feature information is merged,the Softmax classifier is used to identify the action category.The main research work of this paper is as follows:(1)Processing video data using a dense optical flow function.The dense optical flow function is used to calculate the offset of all points on the image to form a dense optical flow field for pixel-level image registration.Then,using two sets of different attribute parameters,two sets of preprocessed data,flow?x and flow?y,are extracted and stored as input of the neural network model.(2)The design is based on the improved 3DCNN human motion recognition model.The improved model is divided into two channels for simultaneous training.The first part of the channel data is first input to the 2DCNN,and the features of the single frame are extracted and amplified,then stacked in time series,input to the 3DCNN,and the feature information is further extracted.The other channel data extracts the feature information from the image optical stream data flow?x and flow?y extracted by the above dense optical flow function as an input of the 3DCNN.Then,the feature information extracted by the two parts of the channel is merged,and the merged feature information is used as the input of the LSTM to further extract the feature information.Finally,the Softmax is used to classify the human motion recognition.(3)KTH data set verification.The experimental platform uses the Py Torch framework,and selects the KTH data set for comparison experiments.It contains 6 types of human motion and 600 AVI video data.In the comparison of similar models,this paper builds a single CNN model,a single 3DCNN model,a single LSTM model and a simple combined model network structure under the same data set.The experimental results show that the proposed method is in the overall recognition rate,single Both the situational dataset and the single-action dataset have improved recognition rates,and the motion recognition effect is better.It laid the foundation for the study of subsequent extended action types and recognition scenarios.