Video Based Mobility Monitoring of Elderly People Using Deep Learning Models

Abstract

In recent years, the number of older people living alone has increased rapidly. Innovative vision systems to remotely assess people's mobility can help healthy, active, and happy aging. In the related literature, the mobility assessment of older people is not yet widespread in clinical practice. In addition, the poor availability of data typically forces the analyses to binary classification, e.g. normal/anomalous behavior, instead of processing exhaustive medical protocols. In this paper, real videos of elderly people performing three mobility tests of a clinical protocol are automatically categorized, emulating the complex evaluation process of expert physiotherapists. Videos acquired using low-cost cameras are initially processed to obtain skeletal information. A proper data augmentation technique is then used to enlarge the dataset variability. Thus, significant features are extracted to generate a set of inputs in the form of time series. Four deep neural network architectures with feedback connections, even aided by a preliminary convolutional layer, are proposed to label the input features in discrete classes or to estimate a continuous mobility score as the result of a regression task. The best results are achieved by the proposed Conv-BiLSTM classifier, which achieves the best accuracy, ranging between 88.12% and 90%. Further comparisons with shallow learning classifiers still prove the superiority of the deep Conv-BiLSTM classifier in assessing people's mobility, since deep networks can evaluate the quality of test executions.

Interaction with a Virtual Coach for Active and Healthy Ageing

“…Ref. [4] describes a study involving the analysis of videos of elderly people performing three mobility tests and how the videos could be automatically categorized using deep neural networks in order to emulate how they are evaluated by expert physiotherapists. It was found that the best results were achieved by a deep Conv-BiLSTM classifier.…”

“…Each LFBL (represented by the next five red blocks) consists of a 2D-CNN, batch normalization, activation, max pooling, and a dropout (20% random drop) layer. The CNN layers within each LFBL are of sizes 64, 64, 128, and 128, and the corresponding max pooling sizes are (2,2), (4,4), (4,4), and (4,4). After the last LFBL, a fully connected layer is used to convert the output of the LFBL into a one-dimensional vector, which is then passed to the LSTM layer (green).…”

A Metadata-Based Approach for Research Discipline Prediction Using Machine Learning Techniques and Distance Metrics

“…A typical approach to compare our research findings to related work is by evaluating the performance of the proposed machine learning models based on the commonly used evaluation metrics such as accuracy, precision, recall, and F1-score on various datasets. This approach is widely adopted in almost every research in the area of machine learning models design [45], [46], [47], [48]. In a similar manner, we compared the proposed approach to related work by applying them to the same dataset and using the same metric to evaluate their performance.…”

Evaluation of the Impact of the Sustainable Development Goals on an Activity Recognition Platform for Healthcare Systems

“…Human activity recognition (HAR) systems monitor activity patterns and intervene when a change in behavior or a critical event has occurred [6][7][8][9][10]. These systems are based on complex architectures that collect data with sensor devices, performing computations and extracting knowledge in order to obtain relevant activity information [11][12][13].…”