Cross dataset non-binary fall detection using a ConvLSTM-attention network
Abbas Shah Syed, Daniel Sierra-Sosa, Anup Kumar, Adel Said Elmaghraby · 2022 IEEE Globecom Workshops (GC Wkshps) · 2022
Fall detection is an important consideration to ambient assisted living research. Falls can result in severe injury and death in the case of the elderly and therefore fall detection systems are pivotal towards the provision of timely care to people experiencing a fall. This paper presents a ConvLSTM (Convolutional Long Short Term Memory)-attention network for the detection of falls considering fall direction and severity. Using data from two publicly available datasets, experiments have been conducted in a non-binary fall versus activity of daily living (ADL) scenario and also a combined ADL recognition and non-binary fall detection scenario. The results for both datasets demonstrate the effectiveness of the proposed method in terms of achieving an average recall of more than 90% in all cases. This work will serve to contribute to the growing utilization of attention mechanisms for time series based classification problems.