| Beginn | 20.10.2015, 10:00 Uhr |
| Ort | TU Braunschweig, Informatikzentrum, Mühlenpfordtstraße 23, 1. OG, Hörsaal M 161 |
| Eingeladen durch | Prof. Dr.-Ing. Marcus A. Magnor |
| Existing context-aware wearables typically directly associate signal patterns to a user’s context, without taking into account previous activities or the person’s location. However, such complementary information may influence the significance of a signal pattern, improve the performance in recognizing context, and enable more fine-grained distinction of activities. The challenge is to achieve such state-aware context recognition from wearables only, i.e., without being dependent on pre-installed infrastructure or the training of complex classifiers. In this talk, we introduce the concept of semantic maps as environment representation for context-aware assistants in daily life. Semantic maps describe the location and state of objects, e. g., open windows, running water taps or closed doors. Such maps can be used to infer higher-level activities of an inhabitant, and to constrain his location within the environment. We discuss a Bayesian filter framework that builds such maps from the data of wearables and in return estimates the person’s position within the map, without the need for any pre-installed localization infrastructure. By continuously adding new information to the semantic map, this proposed framework accumulates knowledge about the place in a self-learning manner. We demonstrate the potential of this tool in real-world settings, where we monitor people during their daily life at home, without having to install any infrastructure in the building. Such a system can be deployed efficiently and without professional support. Furthermore, the introduction of semantic maps as state argument in context-awareness is a major step towards advanced monitoring tools and wearable assistants in medicine, sports and daily-life applications. This talk will presents working prototypes, including the first smartphone-based localization system that robustly tracks people over a long time, using motion data only. | |