Exploring a Deep Learning Approach for Video Analysis Applied to Older Adults Fall Risk

In some medical areas, video activity recognition has been used for patient rehabilitation, evaluating their performance doing some exercises to determine if they are correct or not. In this article we emphasize this approach applied on older adults’ physical activity motivated by the problem caused by falls in this segment of the population. Furthermore, we have developed 8 Deep Learning models to classify different video recordings with the purpose of evaluating and determining how accurately those exercises are executed by people. This article is presented as a first step work, and taking into account this progress, good results were obtained considering the problem of the small number of samples, but addressed including the typical data augmentation techniques. The main results obtained from this work is that the models achieved between 71% and 89% of accuracy, depending on the exercise, and as a conclusion, it allows us to consider this approach to be a valid tool to address the problem of fall risk evaluation in older adults.