This paper investigates the impact of normalizing data acquired from different multimedia sensor devices on the performance of machine-learning-based human fall detection. Specifically, we consider two fall detection datasets (URFD and UP-Fall) and study the impact of eight normalization techniques (min-max, z-score, decimal, sigmoid, tanh, softmax, maximum absolute, and statistical column) on the accuracy and training time of four machine learning classifiers optimized using Grid-Search (namely, support vector machine with radial basis function, k-nearest neighbors, Gaussian Naive Bayes, and decision tree). The conducted experiments confirm that data normalization leads to a significant speed-up in the training of machine learning models and demonstrate which data normalization techniques are the most efficient in terms of accuracy in the context of elderly fall detection.
