Hierarchical Activity Recognition Based on Belief Functions Theory in Body Sensor Networks

Wearable human activity recognition has raised great interest among researchers in the field of Body Sensor Networks (BSNs). However, the classification of human complex activities, especially those similar activities such as walking upstairs and downstairs, is very challenging for the design of efficient activity recognition models. Considering that the hierarchical data structure is helpful to improve the performance of classifiers in effectively detecting the inter-class differences, here, we propose a novel hierarchical activity recognition method based on Belief Functions (BFs) theory. Specifically, in the training stage, we first use Long-Short Term Model (LSTM) to classify the predefined activities and then determine the similarities between each pair of activities through a confusion matrix. Based on this, the activities can be organized into a tree-based hierarchical structure through hierarchical clustering and then the Extreme Learning Machine (ELM) model for each nonleaf node in hierarchical structure can be trained, respectively. In the testing stage, we also propose a novel fast tree-based hierarchical combination rule to fuse the outputs of all trained ELMs. Finally, the proposed hierarchical activity recognition is evaluated by using two widely used UCI datasets: Smartphone and mHealth. The experimental results show that the proposed method is more accurate than the state-of-art methods.