Fast Deep-Learning Based Localization Using Mobile WiFi
"Smart home, healthcare, and intelligent logistics benefit from indoor localization technology. Explore challenges and innovative solutions for real-time performance on mobile devices. Learn about leveraging APs and CSI matrix to improve accuracy and efficiency in location tracking."
Download Presentation
Please find below an Image/Link to download the presentation.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author.If you encounter any issues during the download, it is possible that the publisher has removed the file from their server.
You are allowed to download the files provided on this website for personal or commercial use, subject to the condition that they are used lawfully. All files are the property of their respective owners.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author.
E N D
Presentation Transcript
GLOBECOM 2020 GLOBECOM 2020 MobiFi: Fast Deep-Learning Based Localization Using Mobile WiFi Zhipeng Zhipeng Zhou Zhou1 1, , Jihong Jihong Yu Yu2 2, Zheng Yang , Zheng Yang1 1, Wei Gong , Wei Gong1 1 1 1University University of Science and Technology of of Science and Technology of China China 2 2Beijing Beijing Institute of Technology Institute of Technology 1
Background Background Smart Home Smart Home Health Care Health Care Intelligent Logistics Intelligent Logistics Indoor localization has a promising application 2
State State- -of of- -the the- -art Works art Works Hardware or firmware modification ArrayTrack ArrayTrack (NSDI 13) (NSDI 13) Chronos Chronos (NSDI 16) (NSDI 16) Multiple dimensions SpotFi SpotFi (SIGCOMM 15) (SIGCOMM 15) mD mD- -Track (MobiCom 19) Track (MobiCom 19) Incompatible with commercial device or computationally intensive 3
Challenge Challenge Challenge 1 Challenge 1 But leverage more APs to improve accuracy is not economical. 4
Challenge Challenge PC PC Mobile Device Mobile Device Challenge 2 Challenge 2 Traditional solutions can hardly guarantee real-time performance when deployed on mobile devices. One-step time cost of SpotFi SpotFi 5
Solution Solution -- -- Primer Primer Subcarriers Subcarriers CSI Matrix CSI Matrix ?1,1 ?30,1 ?1,3 ?30,3 6
Solution Solution -- -- SpotFi SpotFi CSI Matrix CSI Matrix Smoothed CSI Matrix Smoothed CSI Matrix CSI Smoothing CSI Smoothing ?1,1 ?30,1 ?1,3 ?30,3 ?1,1 ?30,1 ?1,30 ?30,30 MUSIC MUSIC Location AoA Filtering & Localization Filtering & Localization 7
Solution Solution -- -- Observation Observation CSI Matrix CSI Matrix ?1,1 ?30,1 ?1,3 ?30,3 ???1 MUSIC ???30 Different subcarriers have different localization Different subcarriers have different localization ????????= ???(????? ??????) AoA AoA errors of 30 subcarriers errors of 30 subcarriers for 5 performance! performance! for 5 sample locations sample locations AoA AoA errors of 30 subcarriers errors of 30 subcarriers for 5 for 5 sample locations sample locations 8
Solution Solution We have to select the best subcarrier! We have to select the best subcarrier! No Regular Pattern No Particular Bias We choose to leverage deep learning We choose to leverage deep learning 9
Solution Solution Overview of Overview of MobiFi MobiFi Training Phase Training Phase DL model Input: More training details please refer to our paper! More training details please refer to our paper! Normalized 30*3 CSI Matrix Normalized 30*3 CSI Matrix Best subcarrier index Best subcarrier index groundtruth groundtruth DL model Label: Inference Phase Inference Phase Normalized 30*3 CSI Matrix Normalized 30*3 CSI Matrix DL model Input: Best subcarrier index estimation Best subcarrier index estimation DL model Output: 10
Solution Solution Process of Process of AoA AoA Estimation Estimation CSI of the Best Subcarrier CSI of the Best Subcarrier ? = [??,??,??]?,?? ,??= ? + ? ? ???? ???? ???? ???? ? = ???= 1 ???? = ??(?) ??? ???? 11
Implementation Implementation Open Platform Open Platform Office Office Training Data: Training Data: 28290 28290Locations Locations Testing Data: Testing Data: 8775 Locations 8775 Locations Testing Data: Testing Data: 29775 Locations 29775 Locations Baseline: Baseline: SpotFi SpotFi 12
Implementation Implementation Train Train model on PC with model on PC with Caffe Caffe Deploy trained Deploy trained model model on on NPU NPU Mobile Device: Mobile Device: Raspberry Pi 4B Raspberry Pi 4B Inference Device: Inference Device: Intel Intel Movidius Movidius Neural Computing Neural Computing Stick (NPU) Stick (NPU) 13
Experiment Experiment -- -- Accuracy Accuracy AoA AoA Evaluation Evaluation Location Evaluation Location Evaluation SpotFi: SpotFi: 7.3 7.3 (50%); 15.3 (50%); 15.3 (80 MobiFi MobiFi: : 6.4 6.4 (50 (50%); 14.6 %); 14.6 (80 SpotFi SpotFi: : 1.0m(50 1.0m(50%); MobiFi MobiFi: : 0.9m(50%); 2.7m(80%) 0.9m(50%); 2.7m(80%) (80%) %) %); 2.3m(80%) 2.3m(80%) (80%) %) AoA AoA error performance Comparison error performance Comparison on open platform on open platform MobiFi MobiFi has the same performance as has the same performance as SpotFi Localization performance Comparison Localization performance Comparison on open platform on open platform SpotFi! ! SpotFi SpotFi: : 1.5m(50 1.5m(50%); MobiFi MobiFi: : 1.5m(50%); 4.5m(80%) 1.5m(50%); 4.5m(80%) SpotFi: SpotFi: 8 8.5 .5 (50 MobiFi MobiFi: : 8 8.6 .6 (50%); 21.3 (50%); 21.3 ( (8 80%) %); 4.5m(80%) 4.5m(80%) (50%); 20.9 %); 20.9 (80 (80%) %) 0%) AoA AoA error performance Comparison error performance Comparison on office on office Localization performance Comparison Localization performance Comparison on office on office 14
Experiment Experiment -- -- Time Cost Time Cost PC: PC: SpotFi SpotFi consume consume ~1.1s MobiFi MobiFi consume consume ~0.2s ~1.1s; ; ~0.2s; ; Mobile Device: Mobile Device: SpotFi SpotFi consume consume ~9.5s MobiFi MobiFi consume consume ~1.7s ~9.5s; ; ~1.7s; ; MobiFi MobiFi enables real enables real- -time performance. time performance. 15
Experiment Experiment -- -- Accuracy Accuracy Trained Trained VS VS Untrained Untrained MobiFi MobiFi : : 6.4 6.4 (50 (50%); 14.6 %); 14.6 (80 Untrained: Untrained: 8 8.2 .2 (50%); 1 (50%); 1?.? ( (8 80 0%) Worst Worst: : 18.3 18.3 (50%); (50%); 7 7?.? (80 MobiFi MobiFi : : 0.9m(50 0.9m(50%); Untrained Untrained : : 1.2m(50%); 3.5m(80%) 1.2m(50%); 3.5m(80%) Worst: Worst: 3.8m(50 3.8m(50%); %); 9.9m(80%) (80%) %) %); 2.8m(80%) 2.8m(80%) %) (80%) %) 9.9m(80%) MobiFi MobiFi can be further developed to greatly can be further developed to greatly improve the localization performance. improve the localization performance. MobiFi MobiFi VS VS Oracle Oracle Oracle: Oracle: 0 0.7 .7 (50 MobiFi MobiFi: : 6.4 6.4 (50%); 14.6 (50%); 14.6 (80%) Oracle Oracle : : 0.2m(50 0.2m(50%); MobiFi MobiFi: : 0.9m(50%); 2.7m(80%) 0.9m(50%); 2.7m(80%) (50%); 5.1 %); 5.1 (80 (80%) %) %); 0.6m(80%) 0.6m(80%) (80%) 16
Experiment Experiment MobiFi MobiFi SpotFi SpotFi Component Component Raspberry Pi 4B Raspberry Pi 4B Price Price 199.0 RMB 199.0 RMB 599.0 RMB 599.0 RMB Component Component Price Price PC PC > 2500 RMB > 2500 RMB Intel Intel Movidius Movidius Neural Computing Stick Computing Stick Neural > 2500 RMB > 2500 RMB 798.0 798.0 RMB RMB MobiFi MobiFi is obviously more economical! is obviously more economical! 17
Conclusion Conclusion We present an observation that there exists a We present an observation that there exists a certain subcarrier that can best estimate the location of the target. that can best estimate the location of the target. certain subcarrier We propose a deep learning We propose a deep learning- -based method called automatically automatically select the best select the best subcarrier. based method called MobiFi subcarrier. MobiFi to to We implemented an economical prototype system. We implemented an economical prototype system. 18
Q & A Q & A 19
