1. Gait Identification Using Accelerometer on Mobile Phone
Hoang Minh Thang1,3, Vo Quang Viet1, Nguyen Dinh Thuc2, Deokjai Choi1
1Chonnam National University, South Korea
2Ho Chi Minh University of Science, Vietnam
3Saigon Technology University, Vietnam
Presenter: Hoang Minh Thang

2. Outline
Introduction
Related works
Proposed Methods
Result
Conclusion

3. Outline
Introduction
Related works
Proposed Methods
Result
Conclusion

4. Introduction
The explosion of mobility nowadays is setting a new standard for information technology industry
Mobile devices are commonly used for storage and retrieval of sensitive information (e.g. e-commerce, m-banking, etc.)
Typically poor identification methods
PIN, password
Looking for: Unobtrusive, friendly and implicit authentication mechanism.
Keyword – mobile authentication, gait recognition, accelerometer,
sensor fusion.

5. Introduction
Human gait has been introduced as a particular style and manner of moving human foot. Gait characteristics varies from people to people
Typical gait-based authentication techniques
1. Machine Vision Based
2. Floor Sensor Based
3. Wearable Sensor Based

6. Outline
Introduction
Related works
Proposed Methods
Result
Conclusion

7. Related works
In 2005, H. Ailisto et al. were the first to propose the gait authentication based on wearable accelerometer sensor [2]
S. Terada et al. [3] positioned dedicated sensor on ankle to collect and analyze data
Most of these systems have been implemented with a variety of success rate, but they still have some limitations:
Costly sensors
Need to develop the interface of special sensors
Difficult to implement on the ubiquitous computing environment
In 2009, S. Sprager et al. used an accelerometer on a cellphone positioned at the hip to collect and analyze gait signal [4].
PCA and SVM were adopted on their experiment.

8. Outline
Introduction
Related works
Proposed Methods
Result
Conclusion

9. Proposed Methods Data Acquisition Data Preprocessing Data Segmentation
Time Interpolation
Noise Elimination
Data Segmentation
Gait Cycle Partition
Feature Extraction
Typical Gait Cycles (time domain)
Feature Vector (frequency domain)
Recognition
Template Matching (Dynamic Time Warping)
Classification (SVM)

10. Data Acquisition
The mobile phone (Google HTC Nexus One) was vertically fixed at the pocket location
The sampling rate is approximately 30Hz by using SENSOR_DELAY_FASTEST mode on Android SDK
Fig D coordinate of accelerometer and the trouser pocket position

11. Data Preprocessing Time Interpolation Noise Elimination
Linear Interpolation
Noise Elimination
Collected data contains many noises (e.g., idle orientation shifts, screen taps, bumps on the road…)
Multi-level wavelet decomposition
Daubechies order 6 of level 2

12. Data Preprocessing
Fig. 2. Acquired 3-D acceleration signal when walking
Fig D acceleration after noise reduction

13. Fig. 4. Illustration of a gait cycle
Data Segmentation
Gait Cycle is defined as the time interval between two successive occurrences of one of the repetitive events when walking [5]
Due to the mobile position and ground force reaction, only Z-signal is used to partition gait cycles.
Fig. 4. Illustration of a gait cycle

14. where 𝑑 𝑖 is the 𝑖 𝑡ℎ value in 𝑆(𝑛).
Data Segmentation
The original signal is denoted as S(n).
Extract a set of peaks P from S(n). A data point is called peak if its value is greater than its previous and next one.
Let
Threshold 𝑇 is estimated to determine true peaks using
The peaks which have magnitudes greater than 𝑇 are identified as set of true peaks 𝑅:
P = d i d i > d i+1 ∧ d i > d i−1 } with i ∈ [1…n]
(1)
where 𝑑 𝑖 is the 𝑖 𝑡ℎ value in 𝑆(𝑛).
𝑇 = µ + 𝑘𝜎
(2)
where µ, 𝜎 are mean and standard deviation of all peaks in P respectively ,𝑘 is the user-defined constant
𝑅 = 𝑑 𝑖 ∈𝑃 𝑑 𝑖 ≥ 𝑇 }
(3)

15. Data Segmentation
In our experiment, choosing 𝑘= 1 3 gave the best partition rate
Fig. 5. Illustration of true peaks R and the thresholds T with various k values

16. Fig. 5. Gait cycles and average gait cycle
Feature Extraction
A. Time domain features
Extract average cycles based on partitioned gait cycles [6]
In fact, the walking speed of users could not be constant. Walking speed varies from step to step.
The numbers of data points in every gait cycle are not identical
Dynamic Time Warping (DTW) is used to measure similarity score for avoiding normalizing the length of gait cycles.
Fig. 5. Gait cycles and average gait cycle

17. Feature Extraction A. Time domain features
The magnitude samples of each gait cycle is normalized to values between -1 and +1.
Calculate its distance to every other using DTW
Calculate the average distance 𝑑 𝑖 of one specific cycle to all others.
The average gait cycle 𝐴𝐺𝐶 which has the lowest 𝑑 𝑖 will be considered as the feature template
𝑑𝑡𝑤 𝑖,𝑗 =𝑑𝑡𝑤( 𝑔𝑎𝑖𝑡𝑐𝑦𝑐𝑙𝑒 𝑖 , 𝑔𝑎𝑖𝑡𝑐𝑦𝑐𝑙𝑒 𝑗 )
(4)
where 𝑖 = 1, 2…𝑁−1, 𝑁; 𝑗=1, 2…𝑁 and 𝑁 is the total number of extracted gait cycles
𝐴𝐺𝐶= 𝑔𝑎𝑖𝑡𝑐𝑦𝑐𝑙𝑒 𝑖 | 𝑑 𝑖 = agrmin 1 𝑁− 1 𝑗≠𝑡 𝑁 𝑑𝑡𝑤 𝑡,𝑗
(5)

18. Feature Extraction B. Frequency domain features
First, The length of each gait cycle is normalized to a fixed value
Assume the 𝑖 𝑡ℎ gait cycle can be expressed as
The expected length needs to be normalized is denoted as 𝑇
Calculate the absolute distance 𝐷 between two consecutive data point 𝑆 𝐽 and 𝑆 𝑗+1 .
𝐺𝐶 𝑖 = 𝑆 𝑘 𝑘∈[1…𝐾]},
where 𝑆 𝑘 is the value of 𝑘 𝑡ℎ data point and 𝐾 is the length of this cycle.
𝐷 𝑗 = 𝑆 𝑗 − 𝑆 𝑗 ∀𝑗∈[1…N−1]
(6)

19. Feature Extraction B. Frequency domain features
Determine the position 𝑝 that has the minimum distance 𝐷 𝑚𝑖𝑛 .
The new data point is generated by calculating the average value between the two data point at position 𝑝 and its next one
Two situations:
If 𝐾 < 𝑇, 𝑆 𝑛𝑒𝑤 is added to the position between 𝑆 𝑝 and 𝑆 𝑝+1 .
Otherwise if 𝐾 > 𝑇, we replace 𝑆 𝑝 by 𝑆 𝑛𝑒𝑤 , and remove 𝑆 𝑝+1 .
This process repeats until 𝑇 = 𝐾.
𝑝= 𝑡∈ 1…N−1 ∀𝑗∈[1…N−1]. 𝐷 𝑡 < 𝐷 𝑗 ∧𝑡≠𝑗)
(7)
𝑆 𝑛𝑒𝑤 = 𝑆 𝑝 + 𝑆 𝑝+1 2
(8)

20. Fig. 6. The first 40 FFT coefficients
Feature Extraction
B. Frequency domain features
Second, Fast Fourier Transform (FFT) is calculated using 256-sample window
Split data into windows of eight consecutive gait cycles.
Each window would overlap the previous one by 50% (4 gait cycles)
The first 40 FFT coefficients form a feature vector
Classification using Support Vector Machines (SVM)
Fig. 6. The first 40 FFT coefficients

21. Outline
Introduction
Related works
Proposed Methods
Result
Conclusion

22. Result
Data collected from accelerometer on Google Nexus One phone. The phone position was fixed at the pocket location
A total of 11 volunteers from over 24 year-old participated in data collection.
Each person was asked to walk naturally for an overall of 12 laps with 36 seconds on each lap.
5 of 12 lap data collected randomly were used for training phase and the others 7 lap data were used to predict.
The overall accuracy of 79.1% and 92.7% were achieved respectively to time domain and frequency domain

23. Result Table 1. Confusion matrix of the recognition in time domain
Table 2. Confusion matrix of the classification in frequency domain

24. Outline
Introduction
Related works
Proposed Methods
Result
Conclusion

25. Conclusion
The result indicates that it is possible to implement behavioral biometric-based security on mobile phone in practice
The achieved accuracy is significantly high by an approach on frequency domain.
Limitations
The information retrieved has not been specified
The accuracy of both methods drop significantly when testing on the multiple days dataset
Further work
Validation on a larger dataset
Take a deep look into characteristic of each gait to retrieve features specifically
Analysis gait data in any circumstance(e.g. the effect of time, ground material, phone position, etc.)

26. References
D. J. Fish and J. Nielsen, “Clinical assessment of human gait”, in Journal of Prosthetics and Orthotics 2, April 1993.
H. Ailisto, M. Lindholm, J. Mantyjarvi, E. Vildjounaite and S.M. Makela, “Identifying people from gait pattern with accelerometers” in 2005 Proc. SPIE 5779, Biometric Technology for Human Identification II Conf., USA, 2005.
S. Terada, Y. Enomoto, D. Hanawa and K. Oguchi, “Performance of gait authentication using an acceleration sensor”, in Telecommunication and Signal Processing (TSP), 34th International Conference on, 2011.
S. Sprager and D. Zazula, “A cumulant-based method for gait identification using accelerometer data with Principal Component Analysis and Support Vector Machine”, in Journal WSEAS Transactions on Signal Processing, November 2009.
M. W. Whittle, “Gait analysis an introduction 4th edition”, 2007.
M. O. Derawi, C. Nickel, P. Bours, and C. Busch, “Unobtrusive User-Authentication on Mobile Phones using Biometric Gait Recoginition”, in 6th International Conference on Intelligent Information Hiding and Multimedia Signal Processing, 2010.

27. Thank you
Q & A