Design Article
Wearable sensor devices leverage MEMS motion tracking innovations
Mike Housholder, InvenSense, Inc.
9/28/2012 11:31 AM EDT
Page 4
Activity Recognition1
Activity recognition is one of the key underlying technologies required to make these ideas come to life and this is where motion interface plays a critical role. Activity recognition is the identification of personal activities by analyzing user motion. It uses trained data from the user, then compresses the data into a decision tree format that can be used to identify when the activity occurs in the future. Simply put, activity recognition allows a wearable sensor device to track your activities without the user having to tell the device this he or she is switching from jogging to playing basketball. This automatic “mode switching” simplifies the user interface for the wearable device.
Data from wearable sensing devices is fused and stabilized into body frame (orientation with respect to the sensor), and also world frame (orientation with respect to up, down, left, right, forwards and backwards) accelerations and angular movements. Body and world frame acceleration and rate data in each axis along with the resultant vector magnitude of acceleration and rate are then compressed into ‘features’ capturing 5-10 seconds of data, then computing the average value and standard deviation over that sample window. Additional features can improve performance and accuracy, including frequency-based energy, entropy, or any other independent scalar measure. Energy and entropy require an FFT and thus may be too resource intensive to compute on a small embedded system like a wearable sensor, but may be processed on a smartphone or other device with more computing power. FFT feature results can be helpful for improving accuracy, if needed, but it is not required.
Examples of features include:
During data collection, features are logged and labeled according to the appropriate activity. “Activities” can be any repetitive motion, typically with a dominant frequency below the sample window of 5-10 seconds, such as standing, walking, running, driving, biking, hiking, sitting, using a computer, different swim strokes, etc.
After a full set of data collection (typically hundreds to thousands of data points for each activity), a decision tree classifier can be generated to take a set of features from unidentified user data and logically decide which activity category it best fits into.
From the decision tree, a complete log of an athlete’s activity during the game or a full synopsis of your activity during a busy Saturday can be logged, analyzed and even shared with friends.
In terms of power consumption, it is possible with today’s consumer-grade MotionTracking technology to build thin, light wearable sensor devices that can broadcast data throughout an entire sporting event or record data through an average day. Using creative power management techniques, it may be possible to further extend battery life. Essentially, not every sensor needs to be powered on at all times. Since some sensors expend more power than others, it is possible to track basic motions with the lowest power sensors, then when a change in activity is detected, the other sensors are powered up to perform a 10-15 second activity classification, then they are put back to sleep to optimize battery life.
In summary, the next generation of wearable sensing devices will leverage innovations in MEMS Motion Tracking to change the way we track our personal fitness goals, train our professional athletes and view our sporting events. Application developers, device manufacturers, sports networks, trainers, athletes and the average consumer will find new and creative ways to make use of the valuable data originating from these new devices.
Visit InvenSense.
1. Activity Recognition concept excerpts were referenced from “Activity Recognition and Classification using the InvenSense 9-axis MotionFit SDK” whitepaper by Jonathan Lee, InvenSense.
About the author:
Mike Housholder, senior director of business development, InvenSense, Inc.
See related links:
MEMS-based tennis racquet enhances playing experience
Packaging, low power & custom MEMS process place Invensense at the top in 6-axis motion sensing
Sensor fusion and MEMS for 10-DoF solutions
------------------------
If you found this article to be of interest, visit Medical Designline where you will find the latest and greatest design, technology, product, and news articles with regard to all aspects of clean technologies. And, to register to our weekly newsletter, click here.
Activity Recognition1
Activity recognition is one of the key underlying technologies required to make these ideas come to life and this is where motion interface plays a critical role. Activity recognition is the identification of personal activities by analyzing user motion. It uses trained data from the user, then compresses the data into a decision tree format that can be used to identify when the activity occurs in the future. Simply put, activity recognition allows a wearable sensor device to track your activities without the user having to tell the device this he or she is switching from jogging to playing basketball. This automatic “mode switching” simplifies the user interface for the wearable device.
Data from wearable sensing devices is fused and stabilized into body frame (orientation with respect to the sensor), and also world frame (orientation with respect to up, down, left, right, forwards and backwards) accelerations and angular movements. Body and world frame acceleration and rate data in each axis along with the resultant vector magnitude of acceleration and rate are then compressed into ‘features’ capturing 5-10 seconds of data, then computing the average value and standard deviation over that sample window. Additional features can improve performance and accuracy, including frequency-based energy, entropy, or any other independent scalar measure. Energy and entropy require an FFT and thus may be too resource intensive to compute on a small embedded system like a wearable sensor, but may be processed on a smartphone or other device with more computing power. FFT feature results can be helpful for improving accuracy, if needed, but it is not required.
Examples of features include:
During data collection, features are logged and labeled according to the appropriate activity. “Activities” can be any repetitive motion, typically with a dominant frequency below the sample window of 5-10 seconds, such as standing, walking, running, driving, biking, hiking, sitting, using a computer, different swim strokes, etc.
After a full set of data collection (typically hundreds to thousands of data points for each activity), a decision tree classifier can be generated to take a set of features from unidentified user data and logically decide which activity category it best fits into.
From the decision tree, a complete log of an athlete’s activity during the game or a full synopsis of your activity during a busy Saturday can be logged, analyzed and even shared with friends.
In terms of power consumption, it is possible with today’s consumer-grade MotionTracking technology to build thin, light wearable sensor devices that can broadcast data throughout an entire sporting event or record data through an average day. Using creative power management techniques, it may be possible to further extend battery life. Essentially, not every sensor needs to be powered on at all times. Since some sensors expend more power than others, it is possible to track basic motions with the lowest power sensors, then when a change in activity is detected, the other sensors are powered up to perform a 10-15 second activity classification, then they are put back to sleep to optimize battery life.
In summary, the next generation of wearable sensing devices will leverage innovations in MEMS Motion Tracking to change the way we track our personal fitness goals, train our professional athletes and view our sporting events. Application developers, device manufacturers, sports networks, trainers, athletes and the average consumer will find new and creative ways to make use of the valuable data originating from these new devices.
Visit InvenSense.
1. Activity Recognition concept excerpts were referenced from “Activity Recognition and Classification using the InvenSense 9-axis MotionFit SDK” whitepaper by Jonathan Lee, InvenSense.
About the author:
Mike Housholder, senior director of business development, InvenSense, Inc.
See related links:
MEMS-based tennis racquet enhances playing experience
Packaging, low power & custom MEMS process place Invensense at the top in 6-axis motion sensing
Sensor fusion and MEMS for 10-DoF solutions
------------------------
If you found this article to be of interest, visit Medical Designline where you will find the latest and greatest design, technology, product, and news articles with regard to all aspects of clean technologies. And, to register to our weekly newsletter, click here.
|
|
|
|
|
Navigate to related information