Abstract
•We propose an effective and efficient framework for fall event detection.•We propose a novel MAD-LTP feature descriptor for audio representation.•We develop a fall detection dataset captured under diverse environments to test the robustness of fall detection systems.•Rigorous experimentation was performed on three diverse datasets to investigate the efficacy of the proposed framework.
Fall detection in elder persons may result in long-lasting injury that can have severe consequences for the rest of their lives. Additionally, prolonged delay in emergency treatment after the fall event escalates the chances of mortality. Hence, fall detection at early stage is critical in terms of providing timely aid with little complications and minimize hospitalization expenses. This work aims to provide an effective and efficient healthcare solution to determine the event of fall detection for elderly persons. We aim to address this fall detection problem for elder persons living lonely and encounter issues in case they fall and are unable to call for assistance. In this paper, we present a fall detection framework by proposing a novel feature space mean absolute deviated-local ternary patterns (MAD-LTP) to examine the environmental sounds and used these features to train the BiLSTM for fall events detection. Our proposed MAD-LTP features successfully address the limitations of existing features i.e., non-robust over dynamic pattern detection, brute force optimization, intolerance over non-uniform noise, etc., for fall detection. Performance of our system is evaluated on three diverse datasets i.e., The daily sounds, A3 Fall 2.0, and our in-house developed MSP-UET fall detection dataset. We compared the performance of the proposed framework against the state-of-the-art methods. We obtained an accuracy of 93.5%, 98.29%, and 98% for the daily sounds, A3 Fall 2.0, and our in-house developed MSP-UET fall detection dataset. Experimental findings indicate the reliability of our method for fall event detection.