Fall Detection using Lifting Wavelet Transform and Support Vector Machine

Abstract

Abstract-Frequency domain features of inertial movement enables multi-resolution analysis for fall detection, yet they are computationally intensive. This paper proposes a computationally light frequency domain feature extraction method based on lifting wavelet transform (LWT) which provides computational efficiency suitable for real-time low power devices such as wearable sensors for human fall detection. LWT is combined with support vector machine (SVM) to identify falls from activities of daily living. Performance of the Haar and Biorthogonal 2.2 wavelets were compared with the time domain feature of root-mean square acceleration using a human fall dataset. Results show that the first level detail coefficients features for both Haar and Biorthogonal 2.2 wavelets achieve good overall fall detection accuracy, sensitivity and specificity.

Real-Time AI-Driven Fall Detection Method for Occupational Health and Safety

“…Based on the existing literature, from a model perspective, fall detection can be performed using (i) a threshold-by examining the collected dataset and determining the optimal threshold for a certain feature [26], (ii) traditional machine learning methods (e.g., support vector machines [27], decision-tree-based algorithms, Gaussian mixture models, logistic regression [28]), or (iii) deep learning algorithms (e.g., convolutional neural networks, recurrent neural networks, long short-term memory (LSTM)). Thresholding and machine learning methods heavily rely on crafted features extracted from the data, while deep learning models can provide good performance while operating only on raw acceleration data and skipping the feature handcrafting step [29].…”