Activities Recognition and Fall Detection in Continuous Data Streams Using Radar Sensor

This student paper presents a Quadratic-kernel Support Vector Machine (SVM) based FMCW (Frequency Modulated Continuous Wave) radar system to recognize daily activities and detect fall accidents. Data collected in this work is divided into two different collection modes, namely, snapshots mode (different activities individually collected in isolation) and continuous activity mode (continuous streams of activities collected one after the other). For the continuous activity streams, a sliding window approach with 4s duration and 70% overlapping has achieved 84.7% classification accuracy and subsequent improvement of 2.6% has been proved by using Sequential Forward Selection (SFS) on six participants to identify an optimal feature set. A `tracking' graph has been utilized to verify that the radar system can correctly identify falls as critical events among the other activities.