Zee: Zero-Effort Crowdsourcing for Indoor Localization Anshul Rai, Krishna Kant Chintalapudi, Venkata N. Padmanabhan, Rijurekha Sen Speaker: Huan Yang
Basic Idea Zee is a system that makes the calibration zero-effort, by enabling training data to be crowdsourced without any explicit effort on the part of users. The only site-specific input that Zee depends on is a map showing the pathways and barriers. Zee tracks user walk distance and orientation. Using both of the track data and floor map, Zee can propose a user walk path. Then using the positions along the walk path and RSS correspondingly as training data to build a WiFi database and it will be updated during the time user using it. For any incoming query, Zee applies HORUS or EZ model on the database to estimate the user location.
Example Scenario Inferring a user’s location Backward belief propagation Recording WiFi measurements Using past WiFi measurements to locate subsequent users
Architecture Placement Independent Motion Estimator Counting Steps Estimating Heading Offset Range Augmented Particle Filter WiFi Database
Counting Steps Idle vs Motion: The STD is small when the user is idle. For the motion scenario the STD is very large. The STD is under 0.01g with 99% probability when the user is idle, it is over 0.01g with almost 100% probability when the user is walking.
Counting Steps Repetitive nature of walks: the acceleration pattern for a given user with a particular device placement repeats.
Counting Steps Generates a step occurred event every samples while the user in the WALKING state.
Estimating Heading Offset Range Magnetic offset: usually a characteristic of a given location, depending on the construction and other materials in the vicinity, and typically remains stable with time. Placement offset: usually remains unchanged even when the user takes a turn and changes the direction of walking. Heading offset:
Estimating Heading Offset Range The spectrum of a typical walk: the second harmonic is either completely absent or is extremely weak in the accelerations experienced by the phone in the direction perpendicular to the user’s walk. It is however always present and dominant in the direction parallel to the user’s walk.
Estimating Heading Offset Range Suppose the magnitude of the second harmonic in the Fourier transform along north is and that along west is . Heading offset: or Error estimation: sectors
Augmented Particle Filter As a user continues to walk in an indoor environment, navigating through hallways and turning around corners, the possibilities for the user’s path and location shrink progressively. 4-D particle Particle update
Augmented Particle Filter Forward pass
Augmented Particle Filter Backward pass
WiFi Database HORUS Construct RSS probability distribution Probability of observing RSS at any location Using Bayesian inference to compute and find the maximum likelihood location
WiFi Database EZ Log Distance Path Loss model Distance estimation Standard trilateration
Results Without WiFi database
Results With WiFi database
Results Overall
Conclusion Strength Weakness No need of user participation No need of user initial location Independent of device placement Active learning strategy of database refinement Weakness Floor map needed Particle may not converge For the early queries, the result may not precise
Questions?