A Research Of Neural Network Based Intelligent Recognition Of Human Activity

Motion pose recognition technology has a wide range of applications,including med-ical diagnosis and monitoring,intelligent human-computer interaction,and virtual reality.Compared with vision-based motion pose recognition,inertial sensor-based motion pose recognition uses inertial sensors such as accelerometers to collect target person motion data for analysis and identification.It is not susceptible to environmental interference and does not violate privacy.This is the basic solution adopted in this work.Traditional motion pose recognition uses feature engineering to extract the features of the collected data,and then uses classic machine learning algorithms(such as random forests,decision trees,etc.)to process the features to complete the recognition.However,this method is greatly affected by artificially extracted features,and different features have different effects on different motion poses,resulting in limited recognition accuracy.To improve this limitation,researchers began to use end-to-end neural networks to process data directly,avoiding the limitations of artificial features,and improving the accuracy of recognition to a certain extent.However,the problem with this method is that the accuracy of recognition of similar motion poses(such as going up and down stairs,etc.)is still low,and the accuracy is low when the training data is insufficient.In view of the above problems,this work researches intelligent recognition of motion poses based on neural networks.Not only does it use end-to-end neural network algo-rithms to achieve high recognition accuracy,but also it is difficult to distinguish between similar poses in the existing technology and the amount of personalized data is biased.For low-level problems,a two-level neural network intelligent recognition algorithm and a data augmentation algorithm for motion pose recognition are proposed.The two-level neural network intelligent recognition algorithm effectively recognizes similar actions by splitting the traditional single-level neural network into two-level neural networks.The data augmentation algorithm is based on the human motion pose cycle,which greatly in-creases the artificial data available for training.These two algorithms effectively improve the recognition accuracy.At the same time,in view of the large demand of neural network computing power,in order to meet the real-time requirements,this work also designed a dedicated neural network hardware acceleration module for the algorithm,and completed preliminary im-plementation on FPGA,including convolution,nonlinear The design and implementation of modules such as transformation and pooling have laid the foundation for embedding intelligent recognition of motion poses into real-time portable devices.Aiming at the proposed algorithm and hardware,this work uses the public data set WISDM to complete the experimental evaluation.The recognition accuracy of the algo-rithm reached 87.9 % without the training of personalized data.In the case of hardware supporting two-level neural network,the chip power consumption is only 160 mw.