Continuous Human Motion Recognition Based FMCW Radar And Deep Learning

As the size of radar becomes smaller and the manufacturing cost is lower and lower,radar-based human behavior recognition has been widely used,and it is possible to realize indoor human behavior recognition with only a simple hardware architecture.FMCW radar is a type of sensor which calculate the difference between the emitted signal and the received signal.So,it is suitable for the recognition of indoor human behavior.In this paper,the difference frequency signals of different human actions are obtained based on FMCW radar,and the collected signals are simply processed by signal processing methods,and finally input into the deep learning network for recognition.The specific work is carried out as follows:(1)First,we study the basic principles of FMCW radar,and analyze the echo signals of continuous human actions.We also design actions with practical significance in daily life for common continuous human actions,and build a human continuous action recognition system based on FMCW radar.We collect the echo signals of single and continuous actions to generate a database and extract micro-Doppler features for subsequent deep learning network processing.(2)We use the means of signal processing to analyze the characteristics of different action echo signals from two dimensions of range-Doppler and micro-Doppler timefrequency images to find a suitable method to process the echo signals to make the echo signal contrast between different movements as prominent as possible.We also analyze the timing correlation of echo signals,combine the echo signals with timing significance with deep learning,and identify the types of different actions and their start and end times within a time slice.(3)We analyze the characteristics of continuously changing actions in a time slice,and find that the features of micro-Doppler time-frequency images of continuously changing actions can be abstracted into optical characters.We can encode different actions,and use the idea of optical character recognition(OCR)to segment and recognize continuously changing actions in time.We collect echo signals containing five kinds of actions which contain information of a single action and multiple action combinations and extract micro-Doppler features.We propose a CRNN-based deep learning network,and input the obtained micro-Doppler time-frequency map into the network for training and recognition.The experimental results are divided into training set and test set in 4:1.The sample recognition rate is as high as 96%,and it can also accurately predict the start and end time of unknown actions in terms of timing.(4)We propose a radar continuous action recognition method based on transformer and micro-Doppler features to analyze the echo signal characteristics of continuously transformed human actions.We believe that the recognition of echo signals with timing information has many similarities with natural language processing methods,which essentially translate abstract signals or language into information that humans can understand and understand.First,Inception-based network is used to encode the features of the micro-Doppler signal,and decompose the signal in a time period into multiple shortduration time slices.In this small time slice,the action is considered to be unchanged.Then,the feature-encoded signal is output to the transformer network for recognition.The test results show that the network achieves a recognition accuracy of 95.2% for echo signals containing continuous information of five different actions.