Effect of Reduced Dimensionality on Deep learning for Human Activity Recognition

Inertial sensor is used to measure angular velocity, acceleration etc when attached to a human body part while performing daily or sports activities. Deep learning classifier models such as Long Short-term Memory (LSTM) networks are capable of extracting high dimensional features from signal sequences to represent human activities. However, a trade-off between the computational complexity and the performance of activity recognition system should be maintained if to be used in a low resource device. In this paper, the original signal is segmented into windows and a set of statistical feature measures is extracted from each window. A stacked LSTM network is trained with the sequences of concatenated feature measures instead of the original signal to represent activities. Using feature selection methods: correlation metric as heat map and feature importance, only significant features are selected to train the stacked LSTM network along with conventional classifiers such as support vector machine, random forest and k-nearest neighbor. To evaluate the system performance, a new dataset collection, named Human Physical Activities (HPA) dataset, for sports activities using four inertial sensors is used along with an existing inertial sensor based dataset, HuGaBD. The new dataset consists of nine static and dynamic movements in outdoor setting. The result shows that for all classifiers and both datasets, correlation metric heatmap extracted features result to better performance, except precision values. LSTM, even with limited number of features outperformed, conventional classifiers with 91.1% and 93.4% accuracy for HuGaBD and HPA dataset respectively. Experimental analysis shows that entropy and median feature measure are highly contributing in classification accuracy and theses are essential features for human activity recognition.