An Adaptive Embedded System for Physical Activity Recognition

Basic human activity recognition performed by ubiquitous devices represents one important area of research seeking for systems that are simple to use, reliable, accurate, and of low cost. Each human possesses individual and distinct characteristics in the way that performs physical activities such as walking or running. Therefore, it is important that the activity recognition system can adapt to the person through a mechanism of learning. This paper describes a low cost and adaptable embedded device system for human movement classification using machine learning. Three machine learning schemes were tested to detect five different types of human physical activities: lying down, standing, walking, running and falling, using a training set with 413 instances. The best learning scheme, LogitBoost, obtained 98.8% of average true positive accuracy, performing also very well in the fall detection task. The other two learning schemes, Multilayer perceptron and Simple Logistic, obtained average TP accuracies of, respectively, 97.8% and 98.1%.