Activity Recognition from User-Annotated Acceleration Data Ling ...
1. Activity Recognition from User-Annotated Acceleration Data Ling Bao and Stephen S. Intille Presented by: Hong Lu
2. Key Questions Can low cost wearable sensors be used for robust, real- time recognition of activity? Can training data be acquired from the end user without researcher supervision? Does recognition require user-specific training data? Do more sensors improve recognition?
5. Making researchers to label training examples does not scaleRecognition rates highly depended on how data is collected 95.6% (laboratory data) VS 66.7% (naturalistic settings)
6. Data Collection What’s an accelerometer ? An accelerometer is a device that measures the vibration, or acceleration of motion of a structure.
7. Why Accelerometer ? Many daily activities involve repetitive physical motion of the body or specific postures E.g. Walking, Running, Scrubbing, Vacuuming Low cost, tiny, energy efficient Watch Phone, mp3 player Camera computer Game controller, the wii remote
15. Features Why we need them ? Summarize the data bin Capture useful information What is the desired characteristics of a good feature ? removing irrelevant noise keeping relevant attributes to tell the difference easy to compute ?
16. Features 512 sample windows (6.7s ?), 50% window overlap Features: Mean Energy Frequency-domain entropy Correlation Between x, y accelerometer axes each board Between all pair wise combinations of axes on different boards
18. Training Method 1: User-specific training Train on activity sequence data for each subject Test on obstacle course data for that subject Method 2: Leave-one-subject out training Train on activity sequence and activity data for all subjects but one Test on obstacle course data for left out subject Average for all 20 subjects
19. Results C45 Decision tree wins It shows User-specific training: 71.6 ±7.4 Leave-one-subject-out training: 84.3 ±5.2 Why? Commonalities between people may be more significant than individual variations Larger training set
20. Result Overall, promising Data collected by subjects themselves without supervision Data collected both in and outside of laboratory setting Poorer performance results when… Activities involve less physically characteristic movements , Activities involve little motion or standing still Activities involve similar posture/movement (e.g. watching TV, sitting and relaxing)
21. The dark side The more sensors you placed, the higher accuracy you may achieved, but … cost you look weird hard to deploy more computational horse power
22. Accelerometer Discriminatory Power Tested C4.5 classifier with using subsets of accelerometers: Hip, wrist, arm, ankle, thigh, thigh and wrist, hip and wrist Best single performers: Thigh (-29.5%) Hip (-34.1%) Ankle(-37%)
23. Accelerometer Discriminatory Power With only two accelerometers get good performance: Thigh and wrist (-3.3% compared with all 5) Hip and wrist (-4.8% compared with all 5)
24. Overview The study Activity recognition: 20 household activities Sensors: 5 non-wired accelerometers Data: participants labeled own data Result Good performance with decision tree classifier Subject-specific training data for some activities may not be required Reasonable accuracy can be achieved with only 2 of 5 accelerometers
25. Thank you! The End For some slides, I used content of Emmanuel MunguiaTapia’s presentation