Human Activity Classification Based on Cognitive Doppler Radar to Optimize Carrier Frequency and Sampling Rate Using Reinforcement Learning

We investigate the feasibility of using cognitive radar to increase the accuracy of human activity classification based on micro-Doppler signatures. Micro-Doppler signatures in spectrogram are influenced by various radar parameters, including carrier frequency and sampling rate. In addition, each human body motions have different speeds, which affects the quality of spectrogram. As the quality of spectrogram is the function of radar parameters as well as human motion, there exist certain radar parameters for each activity to capture the micro-Doppler signature with quality, which will ultimately enhance the classification accuracy. This article proposes to use the concept of cognitive radar that changes radar parameters depending on human motion to increase the classification accuracy when deep convolutional neural networks (DCNNs) are employed. Based on the quality of spectrogram measured, we update radar parameters of carrier frequency and sampling rate using reinforcement learning (RL) to maximize classification performance. The term image quality is defined based on the folding rate and definition rate of spectrogram.
${Q}$
-learning was employed to adaptively change the radar parameters given the quality of spectrogram. It is verified against seven human motions that the micro-Doppler classification accuracy can be improved by the operation of cognitive radar.