Fall Detection System Using Novel Median Deviated Ternary Patterns and SVM

In recent years, we have noticed exponential growth in the elderly population of the world due to the advancement in the medical field that necessitates proper care and more attention of elderly people. Accidental falls can be life threatening and can cause severe head trauma, bone fractures, and internal bleedings. Moreover, the most devasting problem of accidental fall incident is that the person remains on the floor for a long time without getting any immediate assistance and response. Research community proposed various fall detection systems but still there exist certain limitations of the existing methods i.e., computational complexity, expensive sensors, unable to wear wearable sensors, and associated privacy issues. To address these issues, we proposed a novel feature descriptor median deviated ternary patterns (MDTP) for audio representation to effectively capture the discriminatory traits of fall and non-fall events. We used the proposed MDTP features to train the support vector machine (SVM) to classify the fall and non-fall incidents. Our proposed method is evaluated against two datasets i.e. A3 fall 2.0 dataset and the MSP-UET fall detection dataset. Our proposed method achieved remarkable accuracy of 98% and 97%, precision of 100% and 96%, recall of 97% and 96%, and F1-score of 98% and 96% on the A3 fall 2.0 and MSP-UET fall detection datasets respectively. Experimental results signify the effectiveness of the proposed system for reliable monitoring of elderly people for fall detection.