부산대학교 도서관

Indoor positioning is an important part of supporting the Internet of Things and location-based services. Crowdsourcing-based magnetic map construction is a key technology to realize wide-area consumer indoor positioning. However, current crowdsourcing-based magnetic map schemes are not suitable for typical indoor scenarios (e.g., shopping malls). The reason is that they ignore the characteristics of crowdsourced data, including short-term trajectory, various pedestrian motion patterns, large-scale data set, and so on. In this article, we propose a novel crowdsourcing-based magnetic map construction method. First, learning-based inertial odometry is used to recover precise user motion trajectories regardless of changes in motion patterns. Then, a keyframe-efficient association method of magnetic time–frequency features is proposed, which is suitable for short-term trajectories of various shapes. Finally, a two-step global estimation optimization is proposed to further eliminate false associations of keyframes and improve the robustness of the method. The feasibility of the proposed method is verified by using a multiuser data set in a typical shopping mall scenario. The proposed method takes a total of 60.8 s to process a 12-h data set (subtrajectories with a duration of 90 s), and the average position error is 1.48 m (with scale correction) and 2.53 m (without scale correction). Compared with the existing crowdsourcing-based magnetic map scheme, the proposed method has been significantly improved in terms of feasibility, accuracy, and efficiency.