Hierarchical Belief K-Nearest Neighbors for Human Activity Recognition

In Wearable Body Sensor Networks (WBSNs), the multi-sensor fusion strategy is widely utilized in Human Activity Recognition (HAR) problems. As the classical decision-level fusion strategy, the Belief Functions (BF) theory often uses a flat structure to fuse the outputs of basic classifiers. However, it is very challenge and difficult to accurately identify confusing activities such as walking upstairs and downstairs based on such flat structure. In this paper, a novel Hierarchical Belief Knearest Neighbors approach (Hie-BKNN) for activity recognition is proposed. First, the data-driven hierarchical tree is constructed by using the spectral clustering strategy. Then, inspired by BF theory and classical K-nearest neighbors, the pair-aware belief masses are calculated for each node and the belief hierarchical tree can be generated. Afterwards, we adopt a novel hierarchical fusion strategy to fuse all involved belief hierarchical trees. Finally, by selecting the probability maximization path on the entire tree, the label of a test sample can be effectively predicted. The experimental results show that our proposed Hie-BKNN approach can precisely recognize daily activities of two different UCI public data sets, i.e., Smartphone and WISDM datasets.