Low-Rank Factorization for Edge Computing in Fall Detection with Wearable Sensors

Machine learning performs fall detection by learning different patterns between people falling and people not falling in the accelerometer signal on wearable sensors. However, the problem of running machine learning on wearable sensors is limited computational resources. This paper proposes using low-rank factorization so that deep learning can run on wearable sensors with limited resources. First, four classes of the fall detection dataset were used with accelerometers from Kaggle: “Free falling,” “sitting down,” “running then falling,” and “running then sitting.” Then, a deep neural network (DNN) training process classifies the four labels. Then, low-rank factorization is applied, whose function is compression on the DNN model—accuracy, precision, recall, and f1-score measures the original model's prediction performance. The number of parameters (params) compares the size of the two models. The test results indicate that the “Running Then Falling” class achieved the highest precision and f1-score, at 0.99 and 0.98, respectively. Following this, the “Running Then Sitting” class demonstrated the highest recall with a value of 0.99. The DNN model's performance in detecting falls based on accelerometer data achieved an overall accuracy of 0.97. Notably, the low-rank DNN model, configured with a rank of 12, proves more efficient than the original model. This efficiency comes with a substantial 53.8% reduction in parameters, achieved without any loss in accuracy.