A Segmentation Technique Based on Standard Deviation in BodySensor Networks

Eric Guenterberg, Hassan Ghasemzadeh, Student Member, IEEE, Roozbeh Jafari, Member, IEEE, andRuzena Bajcsy, Fellow, IEEE

Abstract- Pervasive health monitoring utilizing wearable into actions is a key task. In this paper we introduce awireless sensor nodes can greatly enhance the quality of care segmentation technique based on the notion that regionsindividuals receive. Such systems, while in terms of signal of interest in a data stream correspond to times when theprocessing mostly depend on pattern recognition schemes, readings from a sensor change rapidly during a shortmust operate independently of human interaction for read from A sesor chane apid during ahortextended periods. The lack of a general-purpose pe of t al'es is defineds ti ng whichcomputationally inexpensive algorithm capable of the sensor's values remain relatively constant over ansegmenting sensor readings into discrete actions and non- interval. Empirically this is true for even seeminglyactions has hindered the development of these systems. We continuous actions, such as walking. We present aexamine a segmentation scheme based on standard deviation method for automatically segmenting actions in a sensormetric. We provide experimental verification of the method. stream. Finally, we experimentally verify the

segmentation technique by comparing the results to ourI. INTRODUCTION manual segmentation.

r he development of small sensor platforms fromrelatively inexpensive commercially available parts s I . RELATED WORKhas created interesting new opportunities for data Many researchers have investigated techniques to

collection. The sensor nodes can communicate wirelessly, analyze and record human movement using body sensorfeature limited storage capabilities, and are often networks. Jafari et al. [2] propose a wearable movementdeployed as networks of sensors. These sensor networks monitoring platform. The system consists of lightweighthave been used for a wide variety of sensing tasks from wireless sensor nodes, each equipped with accelerometersindividual health monitoring to large scale environmental and gyroscopes. Sensor readings are separated intosensing: employing sensors measuring environmental primitive actions by manual segmentation in an offlinevariables including temperature, humidity, force, manner. At the next stage, they are classified using a k-acceleration and heartbeat. We are particularly interested NN classifier. In [3], Sherrill et al. introduce a system forin Body Sensor Networks (BSNs) which feature sensor activity monitoring according to a clustering approachnodes placed on different locations on the human body. and hierarchical framework. They utilize unsupervisedThese can be used in hospital or clinical settings as a learing to categorize a dataset and subsequentlyreplacement for tethered sensors or for pervasive lifestyle construct a hierarchy of relationships between clusters.monitoring. For clinical systems, the wireless capabilities They segment the primary recordings by dividing theoffer new freedoms, but are architecturally similar to the signal into fixed time slices with a 660 overlap. Reneveysystems they replace, allowing workstations to manage et al. [4] present a technique for activity classificationmost of the heavy computation. However, BSNs offer the which is used to improve the heart rate estimation in aunprecedented ability to monitor patients in a naturalistic system composed of optical probes and accelerometers.setting for an extended period. The BSNs in these settings Feature vectors are extracted from autocorrelationmust operate for protracted intervals without matrices based on data from fixed size segments. In [5]communicating with a base-station. Therefore the signal and [6], Chambers et al. present classification results ofprocessing and classification must be performed in situ. human gesture for video annotation and retrieval. ToThe limited storage and processing capabilities provide perform segmentation, they look into the stationarymany challenges to architects of these systems. portions of the signal by employing a sliding window.

In patter recognition, segmentation of sensor readings They assume that the acceleration magnitude is very closeto the magnitude of gravity for stationary portions; hence

Manuscript received September 7, 2007.thymdlgaiypunosmgiueofttoarE. Guenterberg, H. Ghasemzadeh, and R. Jafari are with the thymdlgaiypunosmgiueofttoar

Department of Electrical Engineering, University of Texas at Dallas, sga sn asindsrbto.Te esrhRichardson, TX 75083 USA. (emails: {mavpion@student., amount of change in log likelihood of every twoh.ghasemzadeh@student., rjafari@}utdallas.edu). subsequent windows. A sharp change in log likelihood

R. Bajcsy is with the Department of Electrical Engineering andcorsndtoheom neetofaewcin/s.Computer Science, University of California, Berkeley, CA 94720 (email:bajcsy@eecs.berkeley.edu).

978-1-4244-1626-4/07/$25.00 ©)2007 IEEE. 63

Z-axis acceleration

1 000 ..500

-00

200 400 600 800 1000 1200 1400 1 600

Rotation around X-axis (ro1)-~~~. _ m

1 8 _... h. D -.

200 400 600 E0O 1000 1200 1400 1 60

Fig. 2: Manual Segmentation of Sit to Stand

streams back to a laptop for post-processing.Benbasat et al. [7] propose a gesture recognition

platform composed of wireless inertial measurement IV. MANUAL SEGMENTATIONunits. The variance of the data over a fixed window is Manual segmentation can be performed through severalmeasured as the potential metric for activity detection. approaches. Logging the time of each action is a form ofPeriods of activity are determined where the variance is segmentation. Signals can be segmented by synchronizinggreater than a constant threshold value. a video to the signals and dividing based on actions

Although all aforementioned techniques have been observed on screen. It is also possible to segment intosuccessful in providing a system of physical action actions by directly examining the signals and utilizingclassification, they may either use a manual segmentation knowledge of the experimental technique.approach to map original sensor readings to primitive Fig. 2 shows a portion of the sensor signals from theactions (e.g. [2]) or utilize a special purpose auto- eight sensor nodes for the movement "Sit to Stand," andsegmentation scheme. For instance, the fixed time slice highlights the segmentation choices. Each plot in thetechniques used in [3] and [4] provide excellent results diagram contains the signals from all the sensor nodes forfor certain classification schemes; but are not appropriate a given sensor type (such x acceleration). The stand-to-sitfor others. We are investigating more general motion is implicitly present in the data so the subjectsegmentation techniques that are capable of becoming returns to the original position. Therefore, a priori, weintegrated within a wide range of classifiers. The know there will be approximately twenty actions, withtechnique in [5] and [6] does capture entire movements every other one being sit-to-stand. Furthermore, all sit-to-that occur over differing intervals, but it depends on stand movements look similar. Using this information,specific sensor types and a relatively specific posture. and looking for periods of rapid changes in the sensorThe segmentation scheme in [7] uses a hard threshold data, the segments can be distinguished.value and not a moving threshold. Another platform forclassification of human movements is introduced by V. AUTOMATIC SEGMENTATIONMathie et al. [8]. Although they can reliably and easily For automatic segmentation, the morphology of thedistinguish between periods of activity and rest, the desired signals cannot be utilized, as it is not known.measurement device they use is selected very specifically. Furthermore, since it is important to limit inter-nodeIn [9], Li et al. propose a similarity measure for communication communsegmenting and classifying motion streams. Motion . icatinthe. .c sensorsegments are generated by collecting data from primary readis beteen no des sina le.resensor data

stream iS filtered to provide a si nal representing the levelstreams and comparing with predefined reference motions g

Wg

. . . ~~~~~~~~~ofactivity at any given time. We develop a techniqueby using some similarity measures. based on standard deviation.

The "spread" of changes in a data stream for a shortIII. SYSTEMARCHITECTURE ~time interval is represented by the standard deviation overWe are using BSNs for physical movement monitoring, this interval. Due to computational limitations, we use

Our system includes eight sensor nodes arrayed on the standard deviation squared. This is the method proposedbody as shown in Fig. 1. Each sensor node can measure x, by [7]. Formally:y, and z acceleration and angular velocity about the x andy axes. Sampling at 22 Hz, each broadcasts the sensor

64

i+( l2 segmentation. For sit-to-lie and lie-to-sit the ankle and,),X.J_P,U(X,) 0thigh nodes have the same performance (about 10%

C2(X,i) j=i-(n-1)/2 (l) errors), but the wrist nodes have about 60% error.n Generally, over 500/ of the sensors are correctly

where: segmented for each movement, but not always the same

i+(n-1)/2 sensors. To achieve high accuracy for global

E, xi segmentation, we must consider correlations between

yn(X,i) j=i-(n-1)/2 (2) segmentation of individual streams.n

The window is of size n and is centered on point xi,The block diagram for the standard deviation signalactivity is shown in Fig. 3.We observed that during periods of high activity, the

threshold that defines a split between actions is higherthan the threshold that distinguishes an action from a restduring periods of low activity.

The normalizer takes a moving average of the localizedstandard deviations over a larger interval. The localstandard deviation is divided by this average or by aminimum normalization if the average is too low. Thiseffectively introduces a moving threshold based onwhether or not there has been a lot of recent activity.

five -movements- -Fig. 4: Sensor PositionsI Normalizer

Local Standard Deviation g VII. CONCLUSioN AND FUTURE WORK

Moving, In this paper we presented a signal processing modelAverage for segmenting data streams from inertial sensors into

periods of activity and rest. This segmentation is efficient

Figo 3: Standard Deviation Approach enough to be deployed on sensor nodes. Our experimentalFig. ~~~~~~~~~~evaluation highlighted both strengths and weaknesses ofthe technique. The main weakness is a failure to take into

VI. EXPERIMENT account multiple sensor streams for a globalFor the experiment, three subjects performed twenty- segmentation. We have developed a model that combines

five movements ten times each. Each wore the sensor the results from per-stream segmentation to perform a

nodes as illustrated in Fig. 4. We segment the results in global segmentation. This collaborative techniquetwo ways: manually and using the localized standard exhibits significant improvement over per-nodedeviation. We consider the manual segmentation results segmentation.canonical. Errors in the automatic tools are anything thatdiffers from manual segmentation. The exact time value Imp la nCeS

of each segment is not considered important: merelystEFERENFESwhether or not a given region is marked as containing an [1] L. Rabiner, A tutorial on Hidden Markov Models and selected

action or rest. applications in speech recognition. Proceedings of the IEEE,77:257ni86, 1989.

A. Results [2] R. Jafari, W. Li, R. Bajcsy, S. Glaser, S. Sastry, Physical ActivityFor individual sensor data streams, the results closely Monitoring for Assisted Living at Home, In Proceedings of

matched manual segmentation, if the manual International Workshop on Wearable and Implantable BodyI '. 'I. 'I 'I 'I . .1. .1 'I I r-ISensor Net-works (BSN), March 2007, Aachen, Germany.

However, the wrist and ankle nodes experienced almost250/O segmentation errors when compared to manual

65

[4] P Renevey, R Vetter, P Celka, J Krauss, Activity Classification Sign Language in Human Computer Interaction, volume 2298 ofUsing HMM For Improvement of Wrist Located Pulse Detection, LNCS, pp. 9-20. Springer-Verlag, 2002.In Proceedings ofthe Biosignal, 2002.

[8] M. J. Mathie, B. G. Celler, N. H. Lovell, A. C. F. Coster,[5] G. S. Chambers, S. Venkatesh, G. A.W. West, H. H. Bui, Classification of basic daily movements using a triaxial

Segmentation of Intentional Human Gestures for Sports Video accelerometer, In Journal of Medical and Biological EngineeringAnnotation, In Proceedings of the 10th IEEE International and Computing, pp. 679-687, Springer Berlin/HeidelbergMultimedia Modeling Conference (MMMi04), 2004. September, 2004.

[6] G. S. Chambers, S. Venkatesh, G. A.W. West, Automatic Labeling [9] C. Li, S. Q. Zheng, B. Prabhakaran, Segmentation and Recognitionof Sports Video Using Umpire Gesture Recognition, Lecture Notes of Motion Streams by Similarity Search, ACM Tansactions onin Computer Science, Structural, Syntactic, and Statistical Pattern Multimedia Computing, Communications and Applications, Vol.Recognition, pp. 859-867, 2004. 03, No. 03, pp. 1-26, 2007.

[7] A. Y. Benbasat, J. A. Paradiso. An inertial measurementframework for gesture recognition and applications. In Gesture and

66