Analysing transitional activities for a fall using continuous accelerometer sensor data stream

• Main Authors: Abraham, Sonia, James, Rekha K.
• Format: Conference Proceeding
• Language: English
• Subjects: Accelerometers; Accuracy; Cutting equipment; Data transmission; Demand analysis; Fall detection; Gait recognition; Human activity recognition; Human motion; Inertial sensing devices; Recurrent neural networks; Sensors; Smart sensors
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0227424

The prevalence of inertial sensors in smart devices coupled with their cutting edge computational capabilities demand widespread applications in human activity recognition. The continuous data stream from accelerometer sensor facilitates the detection of activities including transitional activities and fall. Nevertheless, the fixed width sliding window technique is inappropriate for detecting transitions in short intervals from a continuous data stream as the discriminant features could go undetected. It may eventually be possible to identify aberrant gait that might lead to a fall by examining the transitional activities. A proposal is made to recognize the transitional activities in between the basic activities from an accelerometer sensor data stream using dynamic window size and RNN-based deep learning models. LSTM, GRU and combination of these RNN variants are used in conjunction with raw accelerometer sensor data for fall detection. The suggested model is assessed using the UCI-HAR and the SisFall datasets. Experimental results showed that Bi-directional LSTM and stacked variants outclassed in recognizing activities and fall using fewer blocks of data. Sedentary activities and fall were recognized with accuracy 94 ± 4% and highly inseparable dynamic activities showed a lower accuracy in comparison.