1. QGesture: Quantifying Gesture Distance and Direction with WiFi Signals
UBICOMP 2018
QGesture: Quantifying Gesture Distance and Direction with WiFi Signals
Nan Yu†
Wei Wang†
Alex X. Liu†‡
Lingtao Kong†
†Nanjing University, ‡Michigan State University

2. Motivation Advantage of WiFi-based Gesture Recognition:
Better coverage
Not need light
Device-free
Motivation
System design
Experiments
Conclusions
2/20

3. Motivation
It is convenient to adjust the volume using the same gesture.
Motivation
System design
Experiments
Conclusions
3/20

4. Motivation
It is important to provide quantitative inputs while interacting with smart home devices.
(a) Microwave oven
(b) Light
(c) TV
(d) Air condition
Motivation
System design
Experiments
Conclusions
4/20

5. Challenges Quantify pushing gesture distance and direction
Reconstruct the phase of CSI measurements
Separate gesture movements from daily activities
Motivation
System design
Experiments
Conclusions
5/20

6. Basic Ideas
QGesture uses COTS WiFi devices to measure the movement distance and direction of human hands
QGesture system overview
Motivation
System design
Experiments
Conclusions
6/20

7. Denoising CSI Signals Noises in CSI Measurements: CFO SFO & PBD
Magnitude Variations
Phase Correction:
Phase of subcarrier 0 for each antenna pair only contains the phase of CFO and the impact of hand movements.
Motivation
System design
Experiments
Conclusions
7/20

8. Denoising CSI Signals Phase Correction: Obervation 1:
Magnitude of static values is more than ten times higher than that of the dynamic value
(a) CSI time series pattern
(b) Variance of different subcarriers
Motivation
System design
Experiments
Conclusions
8/20

9. Denoising CSI Signals Phase Correction: Obervation 2:
Subcarrier with the large magnitude of static values has small change of phase
I/Q phasor representation in two subcarriers
Motivation
System design
Experiments
Conclusions
9/20

10. Denoising CSI Signals Phase Correction:
Pick the CSI phase in subcarrier 0 of one antenna pair that has the largest magnitude of static components to serve as CFO reference
Subtract the CFO in other antenna pairs
Perform a linear regression to remove SFO/PBD
(a) Raw phase
(b) After phase correction
Motivation
System design
Experiments
Conclusions
10/20

11. CSI Signals After Denoising
(a) I/Q waveforms
(b) I/Q trace in complex plane
Pull back 40 cm → path length change 80 cm → phase change around 32π
Phase change direction → movement direction
Motivation
System design
Experiments
Conclusions
11/20

12. 2D Tracking Not the same line of senders and receivers:
Localize the hand by intersecting two ellipses
(a) Different path length changes
(b) Multiple receivers
Motivation
System design
Experiments
Conclusions
12/20

13. Preamble Detection Two gestures: punch and double punch
CSI waveform for the punch gesture
Two gestures: punch and double punch
A logistic regression classifier to detect preamble
An accuracy higher than 92% and a false positive rate lower than 3%
Motivation
System design
Experiments
Conclusions
13/20

14. Implementation Components: A WiFi router Two ThinkPad X200 laptops
A server
(a) 1-D
(b) 2-D
Motivation
System design
Experiments
Conclusions
14/20

15. Results-1D
(a) CDF of pushing 40 cm
(b) Movement direction accuracy
90th percentile measurement error of less than 4 cm at a distance of 1 meter
Detect the pushing direction with accuracy more than 95%
Motivation
System design
Experiments
Conclusions
15/20

16. Results-2D
(a) CDF of absolute direction error
(b) Average measurement error for different directions
Mean absolute direction error of tracking the straight line is degrees
Mean measurement error in distance is 3.75 cm for pushing 40 cm
Motivation
System design
Experiments
Conclusions
16/20

17. Operational Distance
Different distances between the hand and receiver :
Distance measurement accuracy of smaller than 5.5 cm with in distance of 2 meters
Different distance between transmitter and receiver:
Measurement error smaller than 2.7 cm where there is no LOS
Motivation
System design
Experiments
Conclusions
17/20

18. Robustness Different Persons: Different Environments:
Average preamble detection accuracy of 89%
Distance measurement error of 4 cm
Different Environments:
Distance measurement error of less than 5 cm
Different Body Movements:
Average measurement error for single finger movement is about 7 cm
Average measurement error for arm movement
is less than 4 cm
Motivation
System design
Experiments
Conclusions
18/20

19. Conclusions Fine-grained gesture measurements:
Natural device-free interactions
Propose a suite of signal processing techniques and algorithms:
Phase correction algorithm
Implement QGesture on COTS WiFi devices in distance of 2 meters and conduct comprehensive evaluations
Motivation
System design
Experiments
Conclusions
19/20

20. Thank you! Q&A? Q&A Motivation System design Experiments Conclusions
20/20