SenseWit: Pervasive Floorplan Generation Based on Only Inertial Sensing

Mobile crowdsourcing is deemed as a powerful technique to solve traditional problems. But the crowdsourced data from smartphones are generally with low quality, which induce crucial challenges and hurt the applicability of crowdsourcing applications. This paper presents our study to address such challenges in a concrete application, namely floorplan generation. In order to utilize pedestrians' traces for indoor location inference, existing proposals mostly rely on infrastructural references or accurate data sources, which are by nature restricted in terms of applicability and pervasiveness. Our proposal called SenseWit is motivated by the observation that people's behavior offers meaningful clues for location inference. The noise, ambiguity, and behavior diversity contained in the crowdsourced data, however, mean non-trivial challenges in generating high-quality floorplans. We propose 1) a novel concept called Nail to identify featured locations in indoor space and 2) a heuristic pathlet bundling algorithm to progressively discover the internal layouts. We implement SenseWit and conduct real-world experiments in different spaces. Our work offers an efficient technique to obtain high-quality structures (either logical or physical) from low-quality data. We believe it can be generalized to other crowdsourcing applications.