Human Activity and Context Recognition using Lifted Marginal Filtering

Computational causal behavior models can be used for joint human activity recognition and reasoning about the context of an activity, like the location of used objects, which is relevant for assistive systems. Such models are computationally expensive due to the large number of different states that need to be considered. However, the distribution of these states is often highly symmetrical. Lifted Marginal Filtering (LiMa) is an inference algorithm that maintains a suitably factorized state distribution, such that symmetrical factors can be represented compactly. In this paper, we show for the first time the application of LiMa to a complex real-world activity recognition setting based on real IMU data. This is achieved by introducing an operation that prevents the distribution representation to grow indefinitely, by projecting the distribution back to an exchangeable distribution. We show that LiMa needs fewer states to represent the exact filtering distribution, and achieves a higher activity recognition accuracy when only limited resources are available to represent the state distribution.