Human Action Recognition Based On Deep Learning

Human Activity Recognition(HAR)is one of the research topics in the field of pattern recognition and artificial intelligence.It has important research significance and has been applied to many fields at present.The recording forms of human activity information include pictures,videos and sensor data.While the sensor has many advantages,such as low power consumption,small volume and low cost,compared with the computer vision data collection equipment,the use of sensor data to carry out HAR has been widely studied and applied.The traditional machine learning methods have large amount of feature work and need the knowledge of the action domain,while the deep neural network directly deals with the input data of the original sensor,which can greatly reduce the workload.So this paper uses deep neural network to carry out HAR to improve the accuracy of recognition and reduce the computational complexity.First,in view of the problem that the original data is difficult to be directly utilized,the human activity data transformation model is constructed in this paper.The data preprocessing method is used to convert the original data of human behavior into the data format of the subsequent convenient processing,which effectively alleviates the difficult training problems caused by the different noise interference and the different data dimension.Secondly,based on the data conversion model and the depth neural network,the human activity recognition model is built,and the peephole connection algorithm is utilized through experimental analysis in order to solve the problem of insufficient recognition accuracy of the standard LSTM network.Experimental results show that the HP-LSTM network with this algorithm can effectively improve the recognition accuracy,but the training computation cost is increased.Finally,through experiment analysis,use synergistic control algorithm to improve the HPLSTM structure and reduce the parameter calculation of control gates.The experimental results show that the S-LSTM with synergistic control can effectively reduce the calculation consumption of data training,and also enhance the ability to identify the confusing activity of the model.In addition,the data transformation model is extended,the data transformation formula is improved,and the compatibility of sensor data processing is enhanced.