Human activity recognition making use of long short-term memory techniques

The optimisation and validation of a classifiers performance when applied to real world problems is not always effectively shown. In much of the literature describing the application of artificial neural network architectures to Human Activity Recognition (HAR) problems, postural transitions are grouped together and treated as a singular class. This paper proposes, investigates and validates the development of an optimised artificial neural network based on Long-Short Term Memory techniques (LSTM), with repeated cross validation used to validate the performance of the classifier. The results of the optimised LSTM classifier are comparable or better to that of previous research making use of the same dataset, achieving 95% accuracy under repeated 10-fold cross validation using grouped postural transitions. The work in this paper also achieves 94% accuracy under repeated 10-fold cross validation whilst treating each common postural transition as a separate class (and thus providing more context to each activity).