Abstract
With numerous applications in distinct domains, especially healthcare, human activity detection is of utmost significance. The objective of this study is to monitor activities of daily living using the publicly available dataset recorded in nine different geometrical locations for ninety-nine volunteers including young and older adults (65+) using 5.8GHz Frequency Modulated Continuous Wave (FMCW) radar. In this work, we experimented with discrete human activities, for instance, walking, sitting, standing, bending, and drinking, recorded for 10 s and 5 s. To detect the list of activities mentioned above, we obtained the Micro-Doppler signatures through Short-time Fourier transform using MAT-LAB tool and procured the spectrograms as images. The acquired data of the spectrograms are trained, validated, and tested exploiting a state-of-the-art deep learning approach known as Residual Neural Network (ResNet). Moreover, the confusion matrix, model loss, and classification accuracy are used as performance evaluation metrics for the trained ResNet model. The unique skip connection technique of ResNet minimises the overfitting and underfitting issue, consequently resulting accuracy rate up to 91%.