Activity Recognition using Ultra Wide Band Range-Time Scan

Automated detection of Activities of Daily Living (ADL) is gaining importance for its application in smart living, elderly car, healthcare, to name a few. In this paper a novel automated method for detection of human activity is proposed using range time scan data of Ultra Wide Band (UWB) radar. Unlike earlier methods of activity recognition using multi sensor fusion and multi radar setup, we have used a single UWB radar in monostatic mode. In this work, intensity of the reflected UWB signal is used to quantify the amount of scattering leading to formation of range time data matrix. Relevant feature extraction from the range time data via two-dimensional two-directional principal component analysis (2D2D-PCA) is carried out. These features are subsequently used by a random subspace ensemble classifier with k-nearest neighbors algorithm. The proposed method is highly efficient with an average training and testing accuracy of 91.4% and 86.4%, respectively, even on unknown subjects. Moreover the technique resulted in average precision and recall of 0.87 and 0.84, respectively.