1. EWSN08: European Workshop on Wireless Sensor Networks
Activity Recognition from On-Body Sensors: Accuracy-Power Trade-Off by Dynamic Sensor Selection
EWSN08: European Workshop on Wireless Sensor Networks
Piero Zappi, Clemens Lombriser, Thomas Stiefmeier, Elisabetta Farella,Daniel Roggen, Luca Benini, and Gerhard Troster
발표자 : 최재운

2. Contents
Introduction
System Details
Evaluation
Conclusion

3. Contents Introduction System Details Evaluation Conclusion Abstract
Background
Problem Statement
Solution Approach
System Details
Evaluation
Conclusion

4. Abstract In this paper, Authors present Dynamic Sensor Selection.
In order to use efficiently available energy while achieving a desired activity recognition accuracy.
They introduce an activity recognition method.
Activity recognition method
It relies on a meta-classifier that fuses the information of classifiers on individual sensors.
Sensors are selected according to their contribution to classification accuracy.

5. Background Wearable computing
Supporting people by delivering context-aware services
Wearable technology has been used in behavioral modeling, health monitoring systems, information technologies and media development.
Gestures and activities are important
aspect of the user’s context
Small and low-power wireless sensor
nodes are used.
Limited memory and computational power.

6. Background
Wearable computing

7. Problem Statement Wearable computing issue
Trade-off solution is needed!!
High classification accuracy is needed
Large number of sensors distribute over the body.
For high classification accuracy, many sensors should be activated.
Minimize energy use
Sensors have battery limitations.
For enhancing lifetime, minimizing sensor size is needed.

8. Solution Approach Related works about energy use
Adaptive sampling rate and unpredictable duty cycle are representative methods.
In this case, they can not be used to minimize energy use.
Since, user gestures can occur at any time, fixed sensor sampling rate and continuous sensor node operation are needed.
Here, they investigate how to extend network life in an activity recognition system, while maintaining a desired accuracy.

9. Contents Introduction System Details Evaluation Conclusion
System Overview
Metaclassifier for Activity Recognition
Dynamic Sensor Selection
Evaluation
Conclusion

10. System Overview System Overview
System relies on classifier fusion to combine multiple sensor data
Gesture classification is performed on individual nodes using Hidden Markov Models (HMM).
A Naïve Bayes classifier fuses these individual classification results to improve classification accuracy.
System introduce dynamic sensor selection to cope with dynamically changing networks
Most sensor nodes are kept in low power state and they are activated when their contribution is needed to keep the desired accuracy.

11. Metaclassifier for Activity Recognition
This activity recognition algorithm is based on a metaclassifier fusing the contributions from several sensor nodes.

12. Metaclassifier for Activity Recognition
Hidden Markov Models (HMM)
A hidden Markov model (HMM) is a statistical model in which the system being modeled is assumed to be a Markov process with unobserved state.

13. Metaclassifier for Activity Recognition
Features extracted from the sensor data are classified by competing Hidden Markov Models
In this paper, they started with 15 random initial models and select the one that shows best classification accuracy on the training set.

14. Metaclassifier for Activity Recognition
Finally, they fuse the class label using a naïve Bayes technique.

15. Metaclassifier for Activity Recognition
The naïve Bayes classifier
Probabilistic classifier based on the Bayes’ theorem and the hypothesis that the input features are independent
A typical decision rule is to classify an instance as beloning to the class that maximizes the a posteriori probability.
C : Class, Ai : n input attributes
It is hard to compute

16. Metaclassifier for Activity Recognition
The naïve Bayes classifier
Applying the hypothesis of independence and the decision rule they obtain;
The Likelihood is the only parameter that has to be calculated.
Do not need to compute by experiments
Common elements

17. Metaclassifier for Activity Recognition
The naïve Bayes classifier
Defining
tc : the number of training instances for which the C=c and Ai=ai
t : the number of training instances for class c
Some classes c may not have a sample for which Ai=ai.
=> = 0
For this reason, they define as follows;
m : the virtual sample per class added to the training set
p : a priori probability of a certain value for an attribute

18. Dynamic Sensor Selection
Purpose : To achieve a desired classification accuracy while prolonging the system lifetime
To select at run-time the sensors which are combined to perform gesture classification.
The system minimize the number of sensor used.

19. Dynamic Sensor Selection
Example
Activated cluster set of sensors to achieve the desired classification accuracy is first selected ( Cluster Size = D )
All subclusters of size (D-1) must still achieve the desired accuracy

20. Dynamic Sensor Selection
Example
When a node fails, they first test whether the remaining nodes fulfill this condition( sub cluster of size D-1 must achieve desired accuracy)

21. Dynamic Sensor Selection
Example
If not, all the clusters of size D+1 that can be built by adding one idle node to the given cluster are tested.
The one that achieves the best performance is selected

22. Dynamic Sensor Selection
Example
If not, the process is repeated until a cluster that fulfills the condition or no idle nodes are left.
In the latter case the system is not able to achieve the desired performance any more.

23. Contents Introduction System Details Evaluation Conclusion
Evaluation of Activity Recognition Performance
Network Lifetime
Conclusion

24. Evaluation of Activity Recognition Performance

25. Evaluation of Activity Recognition Performance
Purpose : Evaluate the performance of classification as a function of the number of nodes
They perform a set of experiments using 19 nodes placed on the two arms of a tester
They applied their algorithm to clusters of nodes with increasing size (one to 19 nodes).
For each size, they created 200 clusters from randomly selected sensor nodes.
For each cluster size, the average, maximum and minimum classification accuracy is recorded

26. Evaluation of Activity Recognition Performance
Correct classification ratio among random cluster as a function of cluster size

27. Network Lifetime
Dynamic sensor selection scheme vs all sensors
(90% minimum correct classification ratio)

28. Network Lifetime
Dynamic sensor selection scheme vs all sensors
(85% minimum correct classification ratio)

29. Network Lifetime
Dynamic sensor selection scheme vs all sensors
(80% minimum correct classification ratio)

30. Network Lifetime
Network life as a function of the minimum accuracy required

31. Evolution of the network
Network Lifetime
Evolution of the network
On the left, in dark, are the active nodes
On the right, the number of active nodes
A) 80% minimum accuracy. B) 90% minimum accuracy

32. Contents
Introduction
System Details
Evaluation
Conclusion
Pros
Cons

33. Conclusion
Energy aware design aims to extend sensor nodes life by using low power devices and poweraware applications.
Their method minimizes the number of nodes necessary to achieve a given classification ratio.

34. Conclusion 1. System Level Pros
주어진 classification ratio를 만족시키면서 network lifetime을 증가시킬 수 있었음.
전체를 다 사용하는 것 보다 월등히 좋은 lifetime을 가지고 있다는 것을 알 수 있음.
각 노드에서 병렬적으로 datamining을 수행하기 때문에, sensor network 특성에 잘 맞음.
각 센서가 제한된 자원을 가지는 센서네트워크의 특성상 병렬적 처리가 적합함.

35. Conclusion 1. System Level Cons
naïve Bayes classifier 계산시 모든 class의 P(C=c)가 같다는 가정의 신빙성 결여
실험 상 모든 class가 나올 확률이 같다고 하였지만, 사람이 처한 상황 등 기타 조건에 따라 class가 나올 확률이 다를 가능성도 높음.
이러한 확률을 미리 계산하여 계산에 추가를 하였다면, 계산량은 많아지겠지만 정확도를 높일 수 있을 것으로 예상.
naïve Bayes classifier 계산시 (a1, a2, …) 의 independence 가정
각각의 sensor에서 분석한 a1, a2 등이 independence하다는 가정하에 naïve Bayes classifier 를 수행하였음.
한 동작에 대해서 각각의 sensor가 동시에 분석하여 나온 결과물이 independence 하다는 가정은 적합하지 않을 것 같음.

36. Conclusion 1. System Level Cons
Dynamic sensor selection에서 새로운 노드 추가하는 방법에 대한 추가 논의 필요
본 논문에서는 cluster에 새로운 노드를 추가할 시, 모든 조합을 다 맞춰본 후 가장 성능이 좋은 것을 추가하기로 하였음.
이러한 방법은 실시간으로 실행시 overhead가 발생할 수 있기 때문에, 미리 노드별로 priority를 선정하고 이에 맞춰서 새로운 노드 추가 방안 고려.
Network lifetime 늘리는데 더욱 초점을 맞추고자 한다면, idle 노드 중 잔여 배터리가 많은 노드에게 배치하는 방안 고려.

37. Conclusion 2. Literature Level Pros Cons
기본의 data mining 기법 중 신뢰도가 높은 것을 선정하여 classifier로 삼았음.
Cons
타 알고리즘과 비교 부족
본 논문에서는 자신들의 selection 기법과 전체노드가 다 사용되는 방법을 비교.
Network lifetime을 늘리는 것을 더욱 강조하기 위해서는, lifetime을 늘리기 위한 다른 방안들과 직접적이 비교가 더 필요.