1. My Tiny Ping-Pong Helper
He Shan
Shanghai JiaoTong University

2. Background Desperate need of practice on my own
Experience of Wearable computing
Why not a ping-pong helper ??

3. Outline Background Introduction Basic Ideas My Work Result Conclusion
Future Work

4. Outline Background Introduction Basic Ideas My Work Result Conclusion
Future Work

5. Introduction press the button Four basic ways of swing rackets:
(1) forehand push (正手推球)
(2) backhand push (反手推球)
(3) forehand chop (正手搓球)
(4) backhand chop (反手搓球)
press the
button

6. Introduction
Off model
On model

7. Basic Ideas
How to determine whether someone have done the right action or not?
Using “Activity Recognition Chain (ARC)” !
Procedure:
(1) data acquisition and preprocessing
(2) data segmentation
(3) feature extraction (mean, std, energy, etc) and selection
(4) training, classification and identification

8. My Work 1st step : data acquisition Data Type: Tools:
(1) standard data
(2) non-standard data
Tools:
(1) Android smartphone
accelerator & gyroscope
(2) An app
extract data from the phone

9. My Work
Periodic accelerator data
Periodic gyroscope data

10. My Work 2nd step : segmentation
Aim: The data segmentation stage identifies those segments of the preprocessed data streams that are likely to contain information about activities (also often referred to as activity detection or “spotting”)
Approach: Sliding Window
Window moves over the time series data to “extract” data segment
Length : 128 & Overlap: 50%
data
… …
Reduce the computational load

11. My Work 3rd step : feature extraction
Key: use a minimum number of features that still allow the ARC to achieve the desired target performance.
Features:
Minimum, maximum, energy, standard deviation (std), mean, signal magnitude area (SMA= 1 T 𝑡 𝑡+𝑇 𝑥 𝑡 𝑑𝑡 + 𝑡 𝑡+𝑇 𝑦 𝑡 𝑑𝑡 + 𝑡 𝑡+𝑇 𝑧 𝑡 𝑑𝑡 )

12. My Work 4th step : training and classification Method: Training 70%
data
Model
Efficient or not?
Standard
data
Test!!
30%
data
Classify
Sample
data
Fit which class?

13. My Work Support vector machine(SVM) – supervised learning models
Analyze data and recognize patterns, used for classification and regression analysis.
Given a set of training examples, each marked for belonging to one of two categories, an SVM training algorithm builds a model that assigns new examples into one category or the other.
Suppose the distance between the classifier and the point is:
γ i = 𝑦 𝑖 𝛚 𝐓 ∙ 𝐗 𝑖 +b y i =±1
Ensures that the distance between the classifier and the nearest point be as long as possible, that is max 𝛚,b γ
s.t 𝑦 𝑖 𝛚 𝐓 ∙ 𝐗 𝑖 +b >𝛾 𝑖=1,…, 𝑁 𝛚 =1
Let γ= 1 𝛚 , so it equals to 𝑚𝑖𝑛 𝛚,b 𝛚 2
s.t 𝑦 𝑖 𝛚 𝐓 ∙ 𝐗 𝑖 +b > 𝑖=1,…, 𝑁

14. My Work From Lagrange duality, we can transform the question into that
max α 𝑊 𝛼 = 𝑖=1 𝑁 𝛼 𝑖 − 1 2 𝑖,𝑗=1 𝑁 𝑦 𝑖 𝑦 𝑗 𝛼 𝑖 𝛼 𝑗 < 𝑿 𝑖 , 𝑿 𝑗 > 𝑠.𝑡 𝛼 𝑖 ≥0, 𝑖=1,…,𝑁 𝑖=1 𝑁 𝛼 𝑖 𝑦 𝑖 = (∗)
After we figure out α i (SMO algorithm), we have
𝛚= 𝑖=1 𝑁 𝛼 𝑖 𝑦 𝑖 𝑿 𝑖 𝑏=− max i: y i =−1 𝛚 𝐓 𝒙 𝒊 + min i: y i =1 𝛚 𝐓 𝒙 𝒊 2

15. P=N(N-1)/2 combinations
My Work
Support vector machine (SVM)
Problem: 1. Binary classification
2. Unable to classify more than two actions
Solution: One Versus One (OVO) classification
Test data
send
P=N(N-1)/2 combinations
N classes
P=N(N-1)/2
SVMs
vote
Final class
需要简单介绍一下SVM及SVM多分类。
以及 判断是否准确的方法：寻找每一类的重心，然后判断距离；或者计算新来数据与所有内置数据之间的最近的几个距离是否在一定阈值之内。
要有测试结果，用矩阵可以表示测试结果。
加油！ ^^

16. My Work SVM Struct 1 Class 1 1 Class 1 2 4 3
forehand push – 1; backhand push – 2; forehand chop – 3; backhand chop - 4
SVM Struct
1&2 1
1&3
Class 1 1
1&4 2 1
Test data
Class 1 3
2&3 2 4
2&4 4
3&4 3

17. Result
Forehand Push
Backhand
Push
Forehand
Chop
Backhand Chop
Recall 67 1 2
95.71%
Backhand Push 56
93.33%
Forehand Chop 3 62
88.571% 4 45
86.538%
Precision
94.37%
91.8%
89.855%
90%
91.270%
Precision : the percentage of actions which are correctly detected within all possible detected results.
Recall : the percentage of actions which are correctly detected within results that are all determined as that specific action.

18. Conclusion Successfully collect the data using Android smartphone.
Segment the raw data using sliding window.
Extract different features of the data.
Construct classification model.
Successfully classify the test data and tell them whether they are standard data or not.

19. Future Work The construction of smartband.
Acquires sensor data and processes it in the same time.
Data transmission between smartband and cell phone.
Rewrite the Matlab program into Android.

20. Thank you! Q&A