Optimized Dual Stream Spectrum Deconvolution Neural Network for Radar-based Human Activity Recognition

This article proposes an optimized Dual Stream Spectrum Deconvolution Neural Network for Radarbased Human Activity Recognition (DSSDNN-RHAR) for identifying human actions in real time. Range-Doppler maps are transferred into a Dual Stream Spectrum Deconvolution Neural Network using low-cost frequency-modulated continuous wave (FMCW) radar. The system’s power source is an edge device. The findings indicate that the system has a 98.2% accuracy rate and an inference time of 2.95 seconds for five human motions. In an indoor safety application, sounding an alarm when a dangerous action takes place is an important component. As a result, performance during binary classification that is, fall versus non-fall activities is also evaluated, with a 99.8% accuracy rate and a 4% false-negative rate. The edge system’s energy precision ratio is evaluated to ascertain the optimal trade-off between accuracy and computational expense. The system achieves a ratio of energy precision of 1.04, when an optimal proportion would be close to zero.