Enhancing action recognition through simultaneous semantic mapping from body-worn motion sensors.

ABSTRACTLocations and actions are interrelated: some activities tend to occur at specific places, for example a person is more likely to twist his wrist when he is close to a door (to turn the knob). We present an unsupervised fusion method that takes advantage of this characteristic to enhance the recognition of location-related actions (e.g., open, close, switch, etc.). The proposed LocAFusion algorithm acts as a post-processing filter: At run-time, it constructs a semantic map of the environment by tagging action recognitions to Cartesian coordinates. It then uses the accumulated information about a location i) to discriminate between identical actions performed at different places and ii) to correct recognitions that are unlikely, given the other observations at the same location. LocAFusion does not require prior statistics about where activities occur, which allows for seamless deployment to new environments. The fusion approach is agnostic to the sensor modalities and methods used for action recognition and localization. For evaluation, we implemented a fully wearable setup that tracks the user with a foot-mounted motion sensor and the ActionSLAM algorithm. Simultaneously, we recognize hand actions through template matching on the data of a wrist-worn inertial measurement unit. In 10 recordings with 554 performed object interactions, LocAFusion consistently outperformed location-independent action recognition (8--31% increase in F1 score), identified 96% of the objects in the semantic map and overall correctly labeled 82% of the actions in problems with up to 23 classes.