Tack: Learning Towards Contextual and Ephemeral Indoor Localization - PowerPoint PPT Presentation

Tack: Learning Towards Contextual and Ephemeral Indoor Localization With Crowdsourcing Liyao Xiang ECE Dept. Nov. 24, 2017

Indoor Localization ‣ Traditional localization infrastructure is costly. ‣ Most user devices are common smartphones. ‣ We want accurate and cheap indoor localization solutions! 2

Localize by Bluetooth Signals ‣ Bluetooth transmitters (<10$, 50+m range) ‣ Users detect Bluetooth signals for positioning. 3

Localize by Crowdsourcing ‣ Use encountering info to further enhance accuracy. Location errors propagate! 4

Probabilistic Inference Z k,t ‣ User/Bluetooth transmitter locations as clear nodes, and their Z i,t encountering state with other users/ D ij,t transmitters as dark nodes. Z j,t 5

Probabilistic Inference Z k,t ‣ Update the most likely position of the clear nodes repeatedly with Z i,t probabilities conditioned on the D ij,t state of dark nodes. Z j,t 6

Probabilistic Inference D k,t-2 Z k,t-2 D k,t-1 Z k,t-1 Z k,t ‣ Expand the inference D ik,t-2 D ik,t-1 to incorporate each node’s history. D i,t-2 Z i,t-2 D i,t-1 Z i,t-1 Z i,t D ij,t-2 D ij,t-1 D ij,t D j,t-2 D j,t-1 Z j,t-2 Z j,t-1 Z j,t time window = 3 7

Probabilistic Inference forward propagation D k,t-2 Z k,t-2 D k,t-1 Z k,t-1 Z k,t ‣ We not only estimate backward D ik,t-2 D ik,t-1 propagation current locations, but also correct history D i,t-2 Z i,t-2 D i,t-1 Z i,t-1 Z i,t locations. D ij,t-2 D ij,t-1 D ij,t forward propagation ‣ The more information included, the more D j,t-2 D j,t-1 Z j,t-2 Z j,t-1 Z j,t accurate localization. time window = 3 time window = 3 8

Inference algorithm User Accelerometer Step Counter Dead Other Users Magnetometer Reckoning Bluetooth Local transmitters Estimator Encountering Position Estimates With code-level optimization, common smartphones can Architecture support our algorithm. 9

Run on iOS. User Interface 10

Mean Error for All Users in Different Settings. 5 Experiment Setting: 7 beacons, 7 users 5 Users 7 Users 4 Mean error (m) 3 2 1 0 HMM Window = 3 Window = 5 Tested on iPhone 6S. Results 11

Tack: Takeaway ‣ inexpensive ( < 10$ transmitter costs, > 2 years ) ‣ accurate ( 2~4m ) ‣ energy-saving ( 40% less smartphone battery ) ‣ easy to deploy 12

Thank you! Any questions? 13