Using Neural Networks For Indoor Human Activity Recognition with Spatial Location Information

Conventional Human Activity Recognition methods with high quality data information have obtained tremendous achievements in the past decades. In this work, we explore three kinds of neural networks to discuss indoor human activity recognition with spatial location information. First, we calculate the feature vector with the raw spatial location data collected from Ubisense system. Then we build the Artificial neural network (ANN), Convolution neural network (CNN) and Long Short-Term Memory (LSTM) models which are able to recognize six indoor physical activities, and feed feature vectors into these models directly. After that, we evaluate the peak performance and discuss the suitability of each model. Finally, the best performance with spatial location information from different experimenters is obtained by CNN: 87%.