WO2010078186A2

WO2010078186A2 - Default value prediction for sensor ensemble power management

Info

Publication number: WO2010078186A2
Application number: PCT/US2009/069380
Authority: WO
Inventors: Matthai Philipose
Original assignee: Intel Corporation
Priority date: 2008-12-31
Filing date: 2009-12-23
Publication date: 2010-07-08
Also published as: WO2010078186A3

Abstract

A system to converts an arbitrary classifier based on multiple sensors into an equivalent classifier with reduced requirements for powering the associated sensors. When these sensors are battery powered (as part of a mobile system, for instance), this invention can reduce mean time to battery recharge/replacement.

Description

DEFAULT VALUE PREDICTION FOR SENSOR ENSEMBLE POWER

MANAGEMENT

FIELD OF THE INVENTION

[0001] Embodiments of the present invention relate to wearable sensors on various parts of the body and, more particularly, to power management of wearable sensors to extend battery life or recharge intervals.

BACKGROUND INFORMATION

[0002] In many fields, for example in the health fields and the gaming arts, there may be a need for a person to wear a number of sensors on various parts of their body. Such sensors may be used as wearable activity monitors. That is, devices that could monitor and report on the user's daily physical activity patterns. [0003] In a gamming environment, a network of wearable sensors attached, for example, to a user's arms and/or legs may enable a new class of physical game that would allow people to interact with the game. A racing game, for example, could be controlled by how fast somebody can shuffle their feet up and down, or arm and leg movements could control a fighting game. This capability would be similar to systems that use a wireless joystick to control a PC game with the additional benefit that the physical sensors would enable a more realistic gamming experience.

[0004] When left on for long periods, many rich-data sensors see "interesting" data only infrequently. For instance, an RFID reader in a watch may see an RFID tag in less than 0.1 % of its scans because its wearer is not handling objects most of the time. A pendant-style face detector camera may see a person in less than 1 % of frames. In such cases, the sensor produces a default value of "no reading". When the sensors are power hungry, much energy may be spent in processing uninteresting information to produce default values.

[0005] This tends to consume a lot of wasted energy leading to more frequent battery changes or battery recharge intervals than should be necessary.

BRIEF DESCRIPTION OF THE DRAWINGS

[0006] The foregoing and a better understanding of the present invention may become apparent from the following detailed description of arrangements and example embodiments and the claims when read in connection with the accompanying drawings, all forming a part of the disclosure of this invention.

While the foregoing and following written and illustrated disclosure focuses on disclosing arrangements and example embodiments of the invention, it should be clearly understood that the same is by way of illustration and example only and the invention is not limited thereto.

[0007] Figure 1 is a diagram illustrating a user wearing a plurality of wearable body sensors;

[0008] Figure 2 is a flow diagram illustrating the use of default value prediction to evaluate classifier variable according to one embodiment of the invention.

DETAILED DESCRIPTION

[0009] Described is a system to convert an arbitrary classifier based on multiple sensors into an equivalent classifier with reduced requirements for powering the associated sensors. When these sensors are battery powered (as part of a mobile system, for instance), this invention can reduce mean time to battery recharge/replacement.

[0010] Reference throughout this specification to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, the appearances of the phrases "in one embodiment" or "in an embodiment" in various places throughout this specification are not necessarily all referring to the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0011] Figure 1 shows a person 100 wearing a plurality of sensors. The may be camera pendants 102, arm movement sensors 104, RFID tag readers 106, leg movement sensors 108, or a variety of other types. As noted above, power hungry sensors consume much energy processing uninteresting information to produce default values. According to embodiments, other low-power sensors in the ensemble may be able to predict that the high-power sensor is likely to yield a default value. For instance, a light-level sensor with light-level threshold classifier could detect when a video camera is unlikely to detect faces because the surroundings are too dark. Similarly, a wrist-worn accelerometer with body-motion classifier could detect when the wearer's hand motions make him unlikely to pick up RFID-tagged objects.

[0012] Embodiments provide a way to use default-value predictions by low-power sensors to avoid default-value measurements by high-power sensors and thus reduce the overall power consumed by sensor ensembles. Because many high- power sensors produce default values most of the time, and it can cost orders of magnitude less power for low-power sensors to predict these values, this invention can reduce power consumed by sensor-based classification systems significantly. The invention also provides safeguards against the case where the default-value predictor is poor.

[0013] It was not previously recognized that default values for many sensors (that may be part of larger classifiers) are highly predictable using simple sensors, that these values are often overwhelmingly common and that simple predictors can inject these values into any part of a classifier. Embodiments use low-power predictors for default values in classifiers to avoid high-power calculations of those values.

[0014] Embodiments comprise a system that may be run on a computer that that takes the following inputs at each time step:

[0015] 1. A classifier C which represents the relationship between variables v1 ,...,vn, of which some observation variables s1 ,...,sm represent sensors. For instance, C may map variables LightLevel and VideoFrame into the Boolean variable FaceDetected. Of these variables LightLevel and Currentlmage represent sensors.

[0016] 2. An ensemble Si,..., S_m of sensors, where sensor S₁ is mapped to variable S₁. For instance sensors photo-diode and video-camera may be mapped into variables LightLevel and Currentlmage. A sensor may be a combination of software and hardware: for instance, a person-detector sensor may combine video hardware with person-detection software and return just the pixels denoting the person.

[0017] 3. An optional default value predictor P₁ for each sensor S₁ along with threshold probability i. The default value predictor should be trainable using sensor data for default values of sensor S₁, take less power than S₁ to execute, and return the probability pi that S₁WiII generate a default value. Optionally, P₁ may be capable of being informed of a misprediction (to facilitate online re-estimation).

[0018] Given these inputs, embodiments work by evaluating classifier C in each time step. All non-observation variables are treated as usual for classifier C. Any time observation variable s, needs to be processed:

[0019] 1. If S₁ has no predictor associated, set s, to the result of running S₁

[0020] otherwise

[0021] 2. Let P₁ be the associated predictor. Run P₁ to get prediction probability p,.

[0022] If p, is above threshold t,,

[0023] a. Set s, to the default value for S₁.

[0024] b. With probability 1 -p,, set si to the result of running S₁.

[0025] c. If this latter value of S₁ is not the same as the previous one, inform P₁ of the misprediction.

[0026] otherwise

[0027] d. Set s, to the result of running S₁

[0028] Treat s, as usual for an observation variable in C.

[0029] In the above, if S₁ has already been run in a time step, simply re-use its value instead of re-running it in that time step. Also, if recording and adapting to mis-predictions is unimportant, steps 2b and 2c can be omitted. Finally, the default value predictor may reason about multiple default values. In this case, each value will have a threshold probability and 2a will be changed to set variable to the particular default value predicted.

[0030] The work done by this system goes beyond what the original classifier C would do (and other than copying scalar values. The main extra cost is in running the predictor; it is important that default value predictors use much less inexpensive sensors and modest computation. The remaining comparison, sampling and function call operations already have minimal energy requirements.

Also note predictors P₁ may be trained either online using the information from our system in step 2b (if such training can be done fast), or periodicially offline by replaying data with all sensors enabled to get a data stream with correct predictions. [0031] Referring to Figure 2, for variable v 200, in decision box 202 it is determined whether or not v is an observation variable. If it is not, then v is evaluated in the usual way for classifier C 204. If v is an observation variable and associated with a sensor S, then it is determined in box 206 if v has a predictor P associated with it. If no, then v is set to the result of running sensor S 208, a cached value may be used if one exists. If v does have a predictor P, then the system predicts a default value d, and lets the prediction confidence by p in box 210. In decision box 212 it is determined whether p is greater than threshold t for predictor P. If no, then v is set to the result of running sensor S 208, as before. If p is greater than threshold t. If yes, then the system samples the sensor uniformly from real numbers in the range 0...1 at 214. Thus, sensors that are part of a classifier are sampled only when necessary. If after sampling the sample is less than or equal to p in box 216, then v is set to the default value in box 218. If the sample is greater than p, then v is set to the result of running the sensor S in box 220. Thereafter, if v is equal to the default value for P in box 222, the predictor P is informed of a failed classification at box 224.

[0032] In most multimodal classifiers in practice, no systematic effort is made to ensure that sensors that are part of a classifier are sampled only when necessary. In these cases, this invention will clearly result in significant power savings. [0033] The above description of illustrated embodiments of the invention, including what is described in the Abstract, is not intended to be exhaustive or to limit the invention to the precise forms disclosed. While specific embodiments of, and examples for, the invention are described herein for illustrative purposes, various equivalent modifications are possible within the scope of the invention, as those skilled in the relevant art will recognize.

[0034] These modifications can be made to the invention in light of the above detailed description. The terms used in the following claims should not be construed to limit the invention to the specific embodiments disclosed in the specification and the claims. Rather, the scope of the invention is to be determined entirely by the following claims, which are to be construed in accordance with established doctrines of claim interpretation.

Claims

WHAT IS CLAIMED IS: 1. An system, comprising: at least one higher power sensor to be worn anywhere on a user's body, the sensor consuming power when on; a lower power sensor to be worn on the user's body used to predict when the higher power sensor will sense nothing to reduce the power consumed by the higher power sensor.