TrackPuzzle: Efficient registration of unlabeled PDR trajectories for learning indoor route graph

Abstract

This paper proposes to learn the graph of accessible indoor routes by crowd-sourcing inertial sensor data from mobile phones. The inertial sensor data collected by each phone is processed by Pedestrian Dead Reckoning (PDR) algorithm to generate motion trajectories. These motion trajectories are noisy, unlabeled, and are in different coordinate frames. Then the crucial problem to learn the route graph is reduced to firstly find associations among the motion trajectories according to intrinsic features of the trajectories, and then to register the trajectories to generate the route graph, which is called a TrackPuzzle problem. Since finding the longest common subsequence on multiple trajectories is NP-hard, a key contribution of this paper is to convert the multiple trajectory association problem to a trajectory-to-graph association problem. A trajectory is selected as a leader graph and all other trajectories are associated and registered with the leader graph in turn and expand the leader graph gradually. A dynamic programming based Longest Common Sub-sequence on Graph (LCSG) searching method is proposed to solve the trajectory-to-graph association in each step. We further employ a DBSCAN-based outlier detection method to filter out the edges in the association set which are highly inconsistent with others. Then filtered associated edges are registered non-rigidly to expand the leader graph. At last, a hierarchical graph optimization process is applied to the final graphs of different floors to generate multi-floor route graph. Extensive experiments in a synthetic environment and two real-world environments show the more feasibility, efficiency and accuracy of the proposed methods than the state-of-the-art methods.