Debbie Meduna, Dev Rajnarayan, Jim Steele, Sensor Platforms, Inc. - July 22, 2013

Making mobile platforms context-aware is a hot topic in today’s information-rich world. Using the sensors available on these platforms, one can infer the context of the device, its user, and its environment. This can enable new useful applications that make smart devices smarter. It is common to utilize machine learning to determine the underlying meaning in the large amount of sensor data being generated all around us. However, traditional machine learning methods (such as neural networks, polynomial regression, and support vector machines) do not directly lend themselves to application in a power-conscious mobile environment. The necessary techniques for ensuring effective implementation of sensor context are discussed.

Current State of Mobile Context Awareness

Context is the inference of user or environment from sensor measurements on mobile devices. Sensors are in all smart devices such as phones and tablets, as well as many wearable devices such as activity wrist bands [1], shoes [2], headphones [3], and even glasses [4]. Interpreting this sensor data is getting more sophisticated as well. For example: an accelerometer can detect that a user is biking [5]; a front-facing camera can detect whether a user is skeptical of what he is reading [6]; GPS and Wi-Fi positioning can detect which store a user entered [7]; audio can detect whether a user is in a meeting or a night club [8]; and blood pressure sensors can detect whether a person is relaxed or agitated [9].

Several major players have also begun using sensors for mobile context: Samsung's context awareness uses the camera to improve user experience such as eye-tracking to scroll the screen when reading; and Google introduced a new location API at Google I/O 2013 which in part provides an activity recognition result [10].

Tiered Architecture to Enable Always-on
In the above examples, a common technique used to extract information out of sensor data is machine learning. Whether it is neural networks, polynomial regression, support vector machines, expert systems, or another technique, machine learning is typically comprised of resource-hungry algorithms. To make matters worse, context-aware systems are most effective when always-on, but continuous use of most hardware depletes the battery rapidly.

Take Google’s activity recognition as an example. This functionality is available on any Android phone that has downloaded and installed the Google Play services SDK, but there are many complaints about this draining the battery [11]. Furthermore, the detection latency is often observed to be at least 30 seconds (and often much longer). In addition, Google does not specify the expected classification accuracy of their result. These deficiencies need to be rectified to encourage developer adoption of context-aware functionality.

We believe that it is possible to achieve low latency, high accuracy, and low power consumption in sensor-based context results. The key is to adopt a hierarchical approach: lower power systems trigger successively higher power systems when necessary to improve accuracy (see Figure 1). This requires a rigorous design approach to quantify the available accuracy of each sub-system, and empower developers to specify the desired level of accuracy required for their applications. The following sections describe the key components of the tiered architecture developed by Sensor Platforms.

Embedded Change Detector to Minimize AP Wakes

One of the lowest power sensors in a smart phone is the accelerometer, which can be sampled at 50Hz for less than 0.02 mWh of energy. This is a negligible fraction of a typical phone battery which generally has a capacity of more than 5000 mWh. Still, by keeping the main application processor awake just to sample this data uses around 250mWh, completely defeating the purpose of low-power sensing. Over the course of a day, this would significantly drain the device battery. There needs to be a smart always-on trigger system that can continuously analyze incoming sensor data using very low power and memory resources, and wake up a more powerful processor whenever something interesting happens. At Sensor Platforms, we have developed just such a system called a “context change detector” [12].

This architecture enables both low power and low latency. This can be shown by running the change detector at different latencies on various user activities (see Figure 2). By comparing to a simple motion detector (such as that available with the inertial wake functionality of an accelerometer), the severity of the tradeoff between latency and power consumption is mitigated, if not eliminated, by using a smart trigger system. Note that the blue bars are relatively insensitive to latency, indicating that the change detector is responding to changes in the signals, whereas the red bars are inversely proportional to latency, indicating that the simple motion detector is triggering every single time.

Lean Motion Sensor Context Engine to Ensure Low Power

A low-level trigger system, such as the context change detector, greatly reduces how often the sensor data is processed, enabling always-on systems.A further optimization is to abstract out the relevant context information from the raw sensor data, allowing higher-level context algorithms to use these abstracted results rather than the sensor data directly for further inferences. This inertial sensor “context engine” reduces the amount of data that must be processed at higher levels. For example, raw inertial sensor data is typically sampled at 50-100Hz on Android systems, whereas motion context updates typically occur on the order of seconds. Consequently, this middle layer can be thought of as a second-level trigger system for higher level resource-intensive context algorithms.

The ability to abstract meaningful context information out of inertial sensor data may not be readily apparent, but at Sensor Platforms we have successfully demonstrated that this is both possible and accurate. Consider the following example where a device is resting on the table, then picked up and held at the same orientation in hand. The accelerometer signals in the x-, y-, and z-direction for these two contexts seem indistinguishable to the human eye, but not to a properly trained classifier (see Figure 3). Our machine-learning algorithms correctly distinguish off-body and in-hand within a few seconds even for this difficult classification scenario.

Not only is the Sensor Platforms’ FreeMotion Library capable of producing accurate context inferences from inertial sensors, it does so with efficient, lean machine learning algorithms. This leanness is a key aspect of this middle context layer for achieving lower power, and requires thoughtful feature selection. The entirety of the motion-based context engine in the FreeMotion Library can fit on an embedded system.

Usability for Developers Ensures Adoption

With such a low-power mobile context architecture in place, the context information can enable developers through additional virtual sensors. In much the same way that developers can currently request SCREEN_ORIENTATION in Android, rather than interpreting raw accelerometer data, they further should be able to query different context sensors. For motion based context, we see at least three fundamental virtual sensors:

• CARRY: describes how the device is being held or moved relative to the user (e.g. inHand, offBody, inPocket),
• POSTURE: indicates how the user is positioned or is moving relative to the environment (e.g. sitting, walking, lyingDown),
• TRANSPORT: indicates the immediate environment containing the user (e.g. inVehicle, onConveyance).

Algorithms can combine these virtual sensors with each other, and possibly with other system-level sensors (e.g. application data, location data, audio signals, etc.), to generate even more complex context inferences. For example, the accelerometer can detect that a user is biking, but a higher-level context could distinguish between exercising or commuting to work, based on location and time of day.Therefore, the concept of virtual context sensors is one that extends beyond motion-based context to contexts derived from any set of sensors available on the device.

First and foremost, making these new virtual sensors available to developers will enable the next level of context-aware platforms.Second, making these new virtual sensors effective for developers requires the API to report confidence level. Just as location based services provide uncertainty information (best known to the end user as the size of the blue circle on map applications), context results should report reliability and accuracy.

Finally, standardizing these confidence levels requires a methodology to measure confidence on context results. Context accuracies in the industry are currently given as a single pre-determined percentage, e.g. 95% accuracy, but this accuracy only reflects the population for which the algorithm was trained and, more importantly, does not account for real-time fluctuations in the accuracy. For example, the accuracy of a walking context algorithm would likely go down if the user has a mis-step or fumbles the device, and this should be reflected in the sensor output. Determining a consistent standard for context awareness platforms is a difficult but important task. Sensor Platforms is working with industry standards bodies to address this.

Conclusion

The mobile architecture presented here uses machine learning to estimate sensor context, with strong emphasis on low latency and low power.The key enablers of this architecture are the tiered trigger systems developed by Sensor Platforms: the context change detector and the motion context engine, which together minimize AP wakes and reduce triggering of more resource intensive context algorithms.In addition, providing context information in the form of virtual sensors with corresponding confidence levels gives the developer more flexibility.Adoption of the additional virtual sensors introduced here provides an extensible, easy-to-use framework for combining a multitude of sensor information to enable the next level of context awareness apps.

References

[1] http://www.nytimes.com/2012/11/15/technology/personaltech/a-review-of-new-activity-tracking-bands-from-nike-and-jawbone.html?pagewanted=all&_r=0

[2] http://bestrunningwatches.co.uk/reviews/garmin-foot-pod-for-forerunner-review/

[3] http://www.cnet.com/8301-33373_1-57359020/cheap-sensors-enabling-new-smartphone-fitness-gadgets/

[4] http://en.wikipedia.org/wiki/Google_Glass

[5] Mannini, A. and A.M. Sabatini, “Machine Learning Methods for Classifying Human Physical Activity from On-Body Accelerometers,” Sensors 2010, 10, 1154-1175.

[6] http://www.affectiva.com/assets/Measuring-Emotion-Through-Mobile-Esomar.pdf

[7] http://scobleizer.com/2012/07/11/mobile-3-0-arrives-how-qualcom-just-showed-us-the-future-of-the-cell-phone-and-why-iphone-sucks-for-this-new-contextual-age/

[8] http://www.informatik.uni-freiburg.de/~spinello/storkROMAN12.pdf

[9] http://www.cooking-hacks.com/index.php/documentation/tutorials/ehealth-biometric-sensor-platform-arduino-raspberry-pi-medical

[10]http://developer.android.com/reference/com/google/android/gms/location/ActivityRecognitionClient.html

[11] http://forum.xda-developers.com/showthread.php?t=2217982 andhttp://productforums.google.com/forum/#!msg/mobile/0CQTG4PWp24/BrqxhGSeOmkJ

[12] Sensor Platforms Blog article: http://www.sensorplatforms.com/activity-transitions-context-aware-systems/

[13] http://www.anandtech.com/show/5559/qualcomm-snapdragon-s4-krait-performance-preview-msm8960-adreno-225-benchmarks/4

LINK http://www.edn.com/design/sensors/4418673/1/Enabling-low-power-sensor-context-on-mobile-devices-