Wireless sensors for Real-Time Activity Recognition

MITes: MIT Environmental Sensors, A Portable Kit of Wireless Sensors for Naturalistic Data Collection

Despite the promise of pervasive sensing, most researchers today who wish to populate environments such as homes with multi-modal sensors are likely to find this to be a difficult and costly (in time and money) endeavor. Past sensor data collection studies have generally been conducted either in homes that were specially (and laboriously) wired with sensors, in homes wired with a small number of sensors for short periods of time, or in controlled laboratory home simulations. This is perhaps because researchers lack access to affordable, flexible, robust, and easy-to-use tools for the study of behavior and technologies in complex, non-laboratory settings such as homes. During prior work installing sensors in homes, we identified a set of design goals for a portable sensing kit that could be easily retrofitted in existing homes and used in longitudinal pervasive computing experiments. We could not find an existing hardware platform that met these goals and therefore designed and built a sensor system optimized for researcher and subject usability. In this work we introduce MITes: MIT Environmental Sensors, a portable wireless sensor platform that can be used to collect data on people’s activities in non-laboratory settings such as homes.

The MITes platform includes six environmental sensor types: (1) movement using ball, mercury, and reed switches, (2) movement tuned for object-usage detection (using acceleration), (3) light, (4) temperature, (5) proximity, and (6) current consumption. The MITes platform also includes five wearable sensors: (1) accelerometers to acquire body motion information, (2) heart rate, (3) ultra violet radiation exposure, (4) an RFID reader in a wristband form factor, and (5) location beacons. All of these sensors can be used simultaneously, and a single receiver acquires the data, which is sent to a PC or mobile computing device for real-time processing.

Object-usage MITes are small sensors that can be affixed to objects in the environment and broadcast their ID whenever they detect the object has been used. We say that an object as been used whenever it is manipulated either by being directly touched, moved, or held by someone. Object usage MITes can detect open-close states, directional movement, rotation/tilt, and vibration in everyday objects using to a 2-axis accelerometer (ADXL202) sampling at 20Hz with a 5% duty cycle. The accelerometer sensor was chosen because (1) it facilitates the sensor installation since it can be placed onboard eliminating the use of external hanging sensors, and (2) it permits the measurement of motion, tilt, and vibration using a single sensor. The following photographs show examples of the installation of object usage MITes in several everyday objects.

The sensors were designed to be “stick-on and forget” devices that measure characteristics of sudden movement of physical objects in the environment.

Mobile MITes are small sensors that can be worn in different parts of the body to measure 3-axis ±10g or ±2g acceleration. Their small dimension (1.2x1.0 x0.25in), and weight (8.1g including battery) permit them to be embedded or attached to wearable objects such as watches, shoes, or belts without constraining the wearer’s movements. The cost of a single 3-axis prototype is currently only $41.20 (qty of 50), and the average battery life at 200Hz is 31.7 hours at the lowest Tx power from a CR2032 battery. This enables a day of continuous data collection without replacing batteries. The sampling rate could be decreased to extend battery life if required. It is also possible to receive data from multiple (six or more) Mobile MITes working simultaneously by setting them to transmit at different frequency channels.

Mobile MITes are being validated by Stanford Medical School, and have been already used in finding convenient times to interrupt, in the study and correction of balance, in feedback systems for rehabilitation, and in context-aware and activity recognition systems.

Light sensing MITes use a light-to-frequency sensor (TSL235R) to measure ambient light in the range of 320nm to 1050nm. The MITes generic sensor board is used to read the frequency of the sensor output signal and broadcast the intensity value and sensor ID whenever its sensor value change exceeds a specified threshold. The sensor can be set to measure low (0.003-80uW/cm2), or high irradiance (0.08-1kuW/cm2). The sensors have the same form factor as the object usage MITes, cost $21 in quantities of 50, and have a battery life of 620 days at 1Hz.

Temperature sensing MITes use a high accuracy, low power digital output temperature sensor (MAX6677AUT3) to measure temperature in the range of –40C to 125C with an accuracy of ±1.5C at 25C. The MITes generic sensor board is used to read the 1.8KHz PWM sensor output signal, convert it to temperature, and broadcast its value and sensor ID whenever the change exceeds a specified threshold. The sensors have the same form factor as the object usage MITes, cost $20.3 in quantities of 50, and have a battery life of 1309 days at 1Hz.

Electrical current sensing (ECS) MITes are adaptive devices that monitor the current consumption in electrical/electronic appliances over a range of 30mA to 28A. A split core current transformer (CT) is used to passively convert the magnetic fields produced by the current through an appliance’s power cord into a voltage. The voltage is then is full wave rectified, conditioned thorough the use of Op-Amps and input to the MITes ADC. This indirect method (use of CT transformer) is used to avoid costly certification by U.S. regulatory agencies. Finally, two automatically selectable resistors located at the output of the CT, in combination with an adjustable gain stage are used to automatically scale the input voltage to the ADC to obtain a full range of current usage. To select the resistors, and gain values, the ECS MITes require a training period where the maximum current consumed by the appliance being monitored is found automatically.

The current sensor is easily installed by just wrapping its CT round one of the wires in the appliance’s power cord. Their cost is $75.5 dollars in quantities of 50, and their battery life is 14 days in continuous mode (1Hz), and 17 days in the training mode (0.25Hz) running on the CR2032 coin cell.

The following picture shows an example of the type of data generated by the ECS MITes. We can clearly see how the different stages of power consumption of everyday appliances can be identified.

The hearth rate monitor (HRM) MITes are based on the Polar industry standard heart rate wearable chest strap transmitter (WearLink), and receiver. The Polar receiver outputs an active low beat-to-beat pulse signal and the primary job of the MITes generic board is to continuously listen for this incoming heartbeat signal. A voltage level converter IC (MAX3375) is used to convert the 5v BPM signal from the Polar receiver down to the 3v logic level required by the MITes input. The MITes board computes the beats per minute (BPM) value and transmits it to the MITes receiver. Finally, a voltage regulator (TPS79330) is included to regulate the 5v power source from the Polar receiver down to 3V to power the MITes. The HR MITes is powered by a 9V battery.

The MITes transmitter and Polar receiver are compactly packaged to fit in a pill box. The cost of the HRM sensor is $95.5 in quantities of 50, where $49 comes from the cost of the Polar receiver. Its battery life is 2.5 days from a 9V 570mAH Alkaline battery. The sensors are being employed in studying behavior and physical activity in cleaning staff conducted by researchers at Boston University.

UV radiation MITes use a UV sensitive photodiode to provide real-time measurements of the amounts of ultraviolet radiation exposure in the UV index scale. This information can be useful for skin cancer researchers to better understand the sun exposure habits of people. The photodiode used is the is the UVI EryF Photodiode sensor made by Sglux, and has a spectral response close to the CIE’s reference action spectrum. An Op-Amp (TLV2401) current to voltage converter is used to convert the mA current flowing through the photodiode to a voltage that can be read by the MITes ADC. The MITes sensor board periodically samples the voltage, converts it to UV index units, and transmits it wirelessly to the receiver. The sensor is calibrated against the UV indexes reported by the AccuWeather.com weather station website.

The cost of the device is $93.5 in quantities of 50 ($45 from UV diode cost), and its battery life is 2.5 days from a CR2032 coin cell at 1Hz. Currently, the sensors are being deployed in a pilot study headed by Dr. David O'Riordan of the Cancer Research Center of Hawaii and Dr. Karen Glanz of Emory University.

Location beacon MITes provide rough position, and proximity information by positioning a MITes receiver at the location of interest, and making the subjects wear a transmitter beacon. The MITes beacon cycles through its four output power settings as it sends RF packets to the receiver, and the receiver determines distance from the ID, and power settings specified in the packets received. Placing the beacon right next to the receiver will allow packets sent with all four settings to be received. As the beacon is moved away from the receiver, the packets sent with the lowest power setting are lost in succession as the distance becomes greater than their respective transmission ranges.

Since the lowest output power setting in the MITes Tx results in a range larger than room, it was required to further attenuate the output power by adding an external RF attenuator chip (HMC274QS16). It was also necessary to minimize RF leakage in solder and metal tracks by shielding the beacon box with 9 layers of aluminum ducting tape. The ranges obtained by using the four output power levels, and Tx rates of 12Hz in a living setting are 2.5, 10, 15, and 20ft as shown in the following picture. In an open space with line of sight conditions, the ranges are 8. 12.5, 16, and 31 ft. Finally, the cost of the beacon is $45 (qty of 50), and its approximate battery life is 5 days.

Proximity MITes use a passive pyroelectric (PIR) sensor board (KC7783R) to detect motion by measuring the infrared radiation emitted by living organisms (0.15-20µm). The MITes generic board is used to detect the active-low output signal generated by the PIR sensor board whenever motion is detected and to broadcast the sensor ID wirelessly. A 5-3v voltage converter (MAX3377) is needed to convert the output level of the PIR sensor board to the 3v input level required at the external interrupt input of the MITes. Moreover, two voltage regulators (MAX8880) are used to regulate the power voltage provided to the PIR sensor board (5V), and the MITes generic board (3V). The proximity MITes range has a round shape with diameters between 1-12ft obtained by obstructing the sensor field of view by installing lids of various tube lengths in the sensor box. Also, since the sensor is passive, no interference between neighboring units is produced. Finally, the cost of the board is $33.1 in quantities of 50, and its battery life is 47.5 days assuming a sampling rate of 0.5Hz and a 9V 570mA alkaline battery.

The RFID wristband MITes combine a miniature RFID reader and an accelerometer to provide 2-axis acceleration and RFID reading capabilities in a convenient wristband form factor. The RFID reader used is the SkyeTek M1 Mini V3.0, which has a diameter of 25mm and thickness of 2mm. The MITes board controls the reader using ASCII commands issued through the serial port. The reader is powered by a 4.7V prismatic Li-polymer rechargeable battery and the MITes board from the onboard 3V regulator provided by the reader. The antenna is hidden around the wristband and its approximate length corresponds to the diameter of the wrist. Finally, an RF amplifier is used to boost the output signal of the reader and produce a final range of roughly 10 cm. This range is sufficient to allow natural interaction with objects. The RFID tags that can be used with the system are chip tags 13.56MHz compliant using inductive power transfer such as the ISO15693, ISO14443, ISO18000, and EPC. Its cost is $181 with a battery life of approximately 5 hours at a sampling rate of 5Hz and an acceleration sampling of 200Hz.

A keyboard-like multi-switch input sensor that allows for up to 25 inputs in a matrix configuration, and that can be useful for just in time questioning is included in the MITes sensor kit. Its costs is $41 (qty of 80) and its battery life is 45 days in an idle mode or 62 hours while three keys are continuously pressed three times a second and held down for 200 ms using a CR2032 battery.

Tactile feedback MITes use the generic MITes sensor board in combination with a 0.55" diameter x 0.135" thick miniature vibrating motor (from Sanko Electric) to provide 255 levels of tactile feedback. The MITes board is used to listen for incoming ‘turn on’ commands from the MITes receiver. After receiving a turn on command, the MITes board turns on the motor and modulates its vibrating speed using a PWM modulated power signal. The sensor has the same form factor as the object usage MITes, cost ~$20 in quantities of 50.

The following table summarizes some of the tests performed over the different type of MITEs:

MITes Type	Evaluation
Mobile MITes	The testing of the sensors consisted on characterizing the intrinsic noise by recording the acceleration signal generated by leaving three sensor boards sitting undisturbed overnight over a fixed table. The worse case values of the min, max, mean, and standard deviation values for all axes were 776, 804, 789.1, and 3.04 using a ±2g accelerometer.
Heart rate	The performance of the sensors is measured by comparing the BPM values computed against a reference tool manufactured by Polar. The average variation in BPM values is 1.6%, mainly caused by the imprecise analog adjustment dial of the reference tool.
UV exposure	Calibration against the UV indexes reported by the AccuWeather.com weather station website and comparing their final output values. The average variation found was ~1%. The UV orientation sensitivity with respect to the sun was measured by changing the sensor inclination. The average percentage variation with respect to the values at 0° found at 45° was 13.5%
Proximity	The average number of activations when somebody walks (2 activations) and runs (1 activation) underneath the sensor coverage area was computed. The numbers of false positives was also computed by leaving 5 sensors located in an empty room overnight for one week. The number of false positives was zero
Location	The effective range of each of the four RF output power settings in an open space (gymnasium) and in a worse-case environmental noise living laboratory (PlaceLab) were measured. The ranges obtained at 12Hz Tx rates were 8. 12.5, 16, and 31 ft, and 2.5, 10, 15, and 20ft respectively.