1. Human Activity Recognition Using Accelerometer on Smartphones
Symposium 2014
Human Activity Recognition Using Accelerometer on Smartphones
Akram Bayat, Dr. Marc Pomplun, Dr. Duc A.Tran

2. Introduction: 1- Define the problem
Symposium 2014
Introduction: 1- Define the problem

3. Introduction : Motivation
Symposium 2014
Introduction : Motivation
Human activity recognition is an important and challenging research area with many applications in healthcare, smart environments and surveillance and security.

4. Introduction : Computer vision-based techniques
Computer vision-based techniques have been widely used for human activity tracking, but they mostly require infrastructure support.
Using smartphones that can be used under the conditions of daily living is a big advantage.

5. Introduction : Some application
Symposium 2014
Introduction : Some application
In-Building Localization with Smartphones [1]

6. Introduction : some application
Symposium 2014
Introduction : some application
Handling digital entities with the feet[2]
Extract both feet tracking (pose) and movement recognition such as kicking, sliding and rotating.

7. Introduction : Motivation of this study
Symposium 2014
Introduction : Motivation of this study
If people answered honestly to the question, 'What are the reasons why you exercise?', a frequent answer would be to burn calories[3]. 
On a growing scale we use mobile phones for diverse activities in our daily life, such as entertainment, education or information purposes.
Type of Activity
Burning how many calories
Our Motivation is to use smartphone to track the user physical activity and estimate his energy expenditure using 3axis accelerometer

8. Introduction : the Contributions of this work
Symposium 2014
Introduction : the Contributions of this work
Activities of our study include some that not have been widely been studied ( e.g. slow versus fast walking, aerobic dancing)
Less sensory input than existing work, yet able to obtain a comparable accuracy …

9. What we have done in this work : Data Collection
Symposium 2014
What we have done in this work : Data Collection
List of Activities
Running
Slow- Walk
Fast – Walk
Aerobic Dancing
Stairs- Up
Stairs- Down
Acceleration Data collecting
Accelerometer data reader App[4]
Overall 79,573 samples
Frequency 100 Hz

10. Data Collection
Fig. 4. A presentation of tri-axial accelerometer data for a typical subject for different activities

11. Raw Data Preparation 𝐴 𝑥 , 𝐴 𝑦 , 𝐴 𝑧 Digital Low Pass Filter
Accelerometer generates 3-time series along x-axis, y-axis and z-axis:
𝐴 𝑥 , 𝐴 𝑦 , 𝐴 𝑧
𝐴 𝐷𝐶 𝑖 = [𝐴 𝑖 +24× 𝐴 𝐷𝐶 𝑖−1 ] 25
𝐴 𝐴𝐶 =𝐴− 𝐴 𝐷𝐶
Digital Low Pass Filter
Digital High Pass Filter
Gravitational Acceleration : Dc components
Body Acceleration : AC components
𝐴 𝑥_𝐷𝐶
𝐴 𝑦_𝐷𝐶
𝐴 𝑧_𝐷𝐶
𝐴 𝑥_𝐴𝐶
𝐴 𝑦_𝐴𝐶
𝐴 𝑧_𝐴𝐶

12. Creating times series data
Compute the magnitude of acceleration: 𝐴 𝑚
10 time series
𝐴 𝑚
𝐴 𝒙
𝐴 𝒚
𝐴 𝒛
𝐴 𝑥_𝐷𝐶
𝐴 𝑦_𝐷𝐶
𝐴 𝑧_𝐷𝐶
𝐴 𝑥_𝐴𝐶
𝐴 𝑦_𝐴𝐶
𝐴 𝑧_𝐴𝐶

13. Feature extraction : Windowing overlapping
Fig Acceleration plots for the six activities along the z-axis that captures the forward movements.
All six activities exhibit periodic behavior but have distinctive pattern.
We observe that Running and Fast-Walking exhibit very similar pattern.

14. Size of window : windowing overlapping
50% of overlapping
128 samples
128
Good performance

15. Features RMS : Root Mean Squared
MinMax value : difference between maximum and minimum for each window.
Mean : Average on each axis over a time period
Standard deviation
Correlation between different pairs of axes
43 dimensional feature vector

16. Selecting top 5 best classifiers Combination of classifiers
Classification
Individual classifiers
Selecting top 5 best classifiers
Combination of classifiers

17. Individual classifiers
Accuracy
Multilayer perceptron
89.48 %
LibSVM
88.76%
Random Forest
87.55%
LMT
85.89%
Logit Boost
82.54%

19. Confusion Matrix 65 1 3 40 2 84 24 5 4 8 114 47 Dancing Stairs_Down
Slow_walk
Running
Stairs_up
Fast_walk 65 1 3 40 2 84 24 5 4 8
114 47

20. F-measure for each Activity of four best classifiers
Best performance for each activity is obtained
for Multilayer Perceptron

21. Classifier fusion
Combining multiple good classifiers can improve accuracy, efficiency and robustness over single classifiers
The method to combine the classifiers in this work is average of probabilities
Classifiers
Accuracy
Multilayer Perceptron, LogitBoost, LibSVM
91.15%
Multilayer Perceptron, LogitBoost,LibSVM, LMT
90.90%
Multilayer Perceptron, LogitBoost,LibSVM, Random Forest
Multilayer Perceptron, LogitBoost
88.51%
Multilayer Perceptron, LibSVM
88.27%
Multilayer Perceptron, LogitBoost, LibSVM, Random Forest, LMT
81.10%

22. conclusion

23. Thank you

24. Emphasize the Contributions of your Paper
Make sure that you explicitly and succinctly state the contributions made by your
paper. The audience wants to know this. Often it is the only thing that they carry
away from the talk.