Research On Action Recognition Based On Deep Learning And MEMS State Machine

Human action recognition relies on acquiring specific related information such as movement,environment,physiological electrical signals,etc.,and judging human behavior by analyzing the attributes of the action.The action recognition based on Micro-Electro-Mechanical System sensors has been widely used in the fields of intelligent nursing,medical rehabilitation,sports and fitness,etc.This article studies human action recognition from two aspects: theoretical research(deep learning)and engineering design(MEMS finite state machine).This paper builds the following hardware environment.First,a smart phone APP based on the Android system is designed for the collection of deep learning experimental data.In addition,a state machine simulation platform based on the STEVAL sensor module and an experimental platform based on n RF51822+ ultra-low energy Bluetooth are built.MEMS state machine programming and low-level computer design for fast action recognition.This paper uses deep learning algorithms to realize the recognition of human body motion data based on MEMS sensors.By segmenting the feature maps of the convolutional layer and connecting the generated small output maps in a residual-like difference layer,a method based on fusion is constructed,and an IRI-ALSTM dual-channel model based on fusion of internal and external multi-scale features is proposed,using two public data sets UCI HAR,WISDM and pre-processed smart phone APP self-collected data to explore the classification performance of the proposed model,and compare different internal The action recognition rate and calculation load under layered parameters are compared with other classification algorithms to verify the effectiveness of the model proposed in this paper.In addition,a data enhancement method based on the genetic algorithm(EGA-UM)of the unsharp mask combined with the elite retention strategy is proposed,which uses unsharp masks to train Sharpen the collected data,and optimize the sharpening parameters through the genetic algorithm based on the elite retention strategy to generate good extended data that improves the classification performance of the IRI-ALSTM model.Finally,through the study of the finite state machine embedded in the MEMS sensor LIS3 DSH,the state machine simulation platform of the STEVAL sensor is used for real-time data collection and visualization of the fall action,and the state machine programming simulation is realized.The optimized state machine model is directly used in the design of the lower computer of the embedded system of n RF51822+ ultra-low energy Bluetooth,and the fall action is quickly recognized by Bluetooth wireless communication.