You are on page 1of 6

Evaluating C-SVM, CRF and LDA Classification for

Daily Activity Recognition

M’hamed Bilal Abidine*, Belkacem Fergani


Faculty of Electronics and Computer Sciences
USTHB
Algiers, Algeria
abidineb@hotmail.com , bfergani@gmail.com

Abstract—The ability to recognize human activities from time information collected by the sensor data and then
sensed information becomes more attractive to computer used for classification. Feature representations are used to
science researchers due to a demand on a high quality and map the data to another representation space with the
low cost of health care services at anytime and anywhere. intention of making the classification problem easier to
This work compares C-Support Vector Machine (C-SVM), solve. In most cases, a model of classification is used that
Conditional Random Fields (CRF) and Linear Discriminant relates the activity to sensor patterns. The learning of such
Analysis (LDA) for imbalanced dataset to perform models is usually done in a supervised manner (human
automatic recognition of activities in a smart home. This labelling) and requires a large annotated datasets recorded
comparative study offers a guideline for choosing the
in different settings.
appropriate algorithms for automatic recognition of
activities. We conduct several experiments carried out on State of the Art methods used for activities
real world dataset and show that the results obtained with recognition can be divided in two main categories: the so
C-SVM are very promising. C-SVM is able to correct the called generative models and discriminative models [4],
inherent bias to majority class and yields improvement in [5], [6]. The generative methods perform well but require
the class accuracy of activity classification (75.5%) in data modeling, marred by generic optimization criteria
comparison with CRF (70.8%) and LDA (72.4%) methods. and are generally time consuming. Discriminative
methods received the most importance in literature for
Keywords-activity recognition; smart home; sensors their simplicity-model free and good performance with a
network; machine learning
fast prediction speed.
I. INTRODUCTION Motivated by the needs of activity recognition
The number of people over the age of 80 is projected problems, and in order to overcome the generative
to increase more than five times by the year 2050 [1]. models drawbacks and point out the advantage of the
Because the healthcare infrastructure is most likely not discriminative ones, we have compared in this paper,
able to handle this growth, Sensor-based technologies in different discriminative classification methods based on
the home are the key of this problem. Particularly in C-Support Vector Machine (C-SVM) [4], [5],
elderly care, the Activities of Daily Living (ADL) like Conditional Random Fields (CRF) [7] and Linear
eating, bathing, dressing and toileting, are used to assess Discriminant Analysis (LDA) [4], [8]. SVM is a powerful
the cognitive and physical capabilities of an elderly technique, which can predict not only for the labelled
person [2]. data, but also for the unlabelled data. It works well for
both linear and non linearly separable datasets. We utilize
The ability to identify the behaviour of people in a in this paper the current standard soft-margin C-SVM [5].
smart home with real-time response is at the core of The CRF should be ideal for modelling the human
ubiquitous computing applications. Smart systems are activities. CRF are thus especially suitable for
equipped with sensor networks able to automatically classification tasks with complex and overlapped
recognize activities about the occupants and assist features. However, CRF requires labelled data. The LDA
humans in an intelligent manner [3]. The sensors can be is a frequently applied classification method for
attached to either an actor under observation or objects predicting labelled data owing to its simplicity; however,
that constitute the environment. it can not predict unlabelled data. Furthermore, LDA may
Sensor data collected are often analyzed using data not work for non linearly separable dataset.
mining and machine learning techniques to build activity The remainder of this paper is organized as follows.
models and perform further means of pattern recognition Section ΙΙ presents related works in activity recognition.
[4], [5]. Recognizing a predefined set of activities is a Section ΙΙΙ describes C-SVM, CRF and LDA algorithms.
classification task : features are extracted from the space - Section IV presents and discusses the experimental

978-1-4673-1520-3/12/$31.00 ©2012 IEEE


comparison based on the above algorithms. Finally, 1
Section V concludes by summarizing our findings and max  im1  i   im1  mj1  i j yi y j K ( xi , x j ) (2)
i 2
outlines future directions.
Subject to im1  i yi  0
II. RELATED WORK 0   i  C , if yi  1
Activity recognition has been carried out on different
types of data. We can cite for example, state change 0   i  C , if yi  1
sensors [9], motion detectors [10], cameras [11], Where  i  0 are Lagrange multipliers. The training
accelerometers [12], RFID tags and sensors [3], [13],
electrical signatures [14], GPS [15]. These technologies samples for which Lagrangian multipliers are not zero are
include different levels of complexity and technological called support vectors.  (.) is a non-linear function
challenges in terms of price, intrusiveness and the type of which maps the input space into a feature space.
data they output [3], [16].
In a soft margin support vector machine, we need to
Activity recognition supervised models are based on select a kernel function K(.,.) and the regularization
large annotated datasets. Annotation has been performed  
parameters C and C in each binary classifier, to
in many different approaches. The least interfering way is construct a classifier for multiple classes.
to use cameras [10]. Other examples are self-reporting
approaches, keeping an activity diary on paper or using a The kernel function can be of various types [4]. The
PDA [17]. RBF kernel has less numerical difficulties; its
computation is not as complex as the other kernel’s. So
Methods used for recognizing activities can be we choose the radial basis kernel function (RBF) :
divided in two categories: generative models (e.g.,
K ( x , y )  exp   12 x  y  where  is the width
2
Hidden Markov Models (HMM) [3], Gaussian Mixture
Models (GMM) [18], Naive Bayes Classifier (NBC) [9])  2 
and discriminative models (e.g., Support Vector parameter. The construction of such functions is
Machines (SVM) [19], k-Nearest Neighbor (k-NN) [20]). described by the Mercer conditions [21].
In this paper, the performances of three discriminative  
C and C are regularization parameters for positive
models named C-SVM, CRF and LDA are assessed and and negative classes, respectively. They are used to
compared in terms of classification accuracy and class control the trade-off between margin and training error
average accuracy [3]. The main purpose of this work is to represented by slack variables ξi in order to avoid the
provide a fair and extensive comparison of these problem of overfitting [4]. Some authors [5], [22] have
commonly employed classification methods using the proposed adjusting different penalty parameters to
same dataset and feature representation. different class, the SVM algorithm effectively improves
the low classification accuracy caused by imbalanced
III. METHODS FOR ACTIVITY RECOGNITION samples (e.g. many positive and few negative). This can

be done, by increasing the tradeoff C associated with the
A. C-Support Vector Machines (C-SVM)
negative class. This is exactly what Veropoulos et al [22]
Support Vector Machines (SVM) is based on propose in their paper.
statistical learning theory developed by Vapnik [5].
However, we utilize the current standard soft-margin C- Solving equation (2) for  gives a decision function
SVM [5]. SVM classifies data by determining a set of for classifying a test point x  R n [4]
support vectors, which are members of the set of training
msv
f ( x )  sgn    i y i K ( x, xi )  b 
inputs that outline a hyperplane into a higher dimensional
(3)
space (feature space).  i 1 
For a two class problem, we assume that we have a Whose decision boundary is a hyperplane and
m translates to nonlinear boundaries in the original space
training set xi , yi  where x  R and yi are class
n
i 1 n
and m sv is the number of support vectors xi  R .
labels either 1 or -1. The primal formulation of the soft-
2 In this study, a software package LIBSVM [23] was
margin in SVM maximizes margin and used to implement the multiclass classifier algorithms. It
K ( w, w)
uses the one-versus-one method [5]. This method consists
minimizes the training error ξi simultaneously by solving in constructing N ( N  1) / 2 classifiers and each one is
the following optimization problem
trained on data from two classes. After all
1   N ( N  1) / 2 classifiers are constructed, a voting strategy
min K (w, w)  C  i  C  i
w,b, 2 yi 1 yi 1 is used for testing. Using this strategy, each binary
(1)
T classification is considered to be a voting, and then the
subject to yi (w  ( xi )  b)  1  i , i  0, i  1,..., m
point is predicted in the class with the largest vote. So it
The dual formulation of the soft margin SVM can be is also called the ‘‘Max Wins’’ strategy. See Figure 1, in
reformulated as: case N  3.

Providing training sets with full labels X , Y , the
m m

parameters of a CRF are learned by maximizing the
conditional probability distribution P (Y | X ) which
belongs to the family of exponential distributions as in
[7]. Potential function takes a specific form
exp( n  f n ( yt , yt 1 , X )) , where n is group of real
weights. These weights are the parameters we want to
Figure 1. OvO-SVM classification of three linearly separable classes find when learning with the model. Model parameters can
( N  3 ) [19]. be learned using an iterative gradient method.
C. Linear Discriminant Analysis (LDA)
B. Conditional Random Fields (CRF) Discriminant analysis approach is well known to learn
A Linear-chain CRF [3], [7] is a popular discriminative feature transformation in the statistical
discriminative temporal probabilistic model. CRF, pattern recognition literature [4], [8] and can be easily
directly represents conditional distributions of the hidden extended to multi-class cases. It is used to reduce the
labels yi given all the observable variable xi at each time dimensionality of data and for classification purposes.
step, As we can see in Figure 2. This assumes that they have a set of observations with
known class labels.
Given an n-dimensional space set of observations:
  i n
Di  x1i ,..., xmi i ( x j  R ) from class Ci (i  1,..., N , N
is the number of classes), we assume that each of the
class probability density functions can be modeled as
normal distribution. We define the prior
probabilities p (Ci ) , means mi , and covariance matrices
Σi of each class as:
Figure 2. The graphical representation of a linear-chain CRF. Grey mi 1 mi
nodes represent observable variables, while white the nodes represent
p (C i )  , mi   xi (8)
hidden ones [3].  mj mi i 1
j

Let X = xi the observed sequence and Y = yi the label 1 mi


T
sequence with a pair K = (Xk; Yk) a set of cliques [7]. The Σi   ( xi  m )( xi  m )
mi i 1
conditional probability of the classes conditioned on the
observed sequence is given by Where mi is the number of patterns in class Ci . With
1 LDA all classes are assumed to have the same covariance
P (Y | X )    ( X k , Yk ) (4) matrices Σi ,…,ΣN, in (8). We assign the new feature
Z  ( X ) kK ( X ,Y )
vector that is to be classified x to Ci with the linear
Where θ is the model parameter to be estimated, 
discriminant function d i . This function is obtained by
is the potential function of clique k and Z ( X ) is the simplifying the quadratic discriminant rule [4]
normalisation function. 1 T 1 T 1
d i ( x)  log( p (Ci ))  mi SW mi  x SW mi (9)
Z ( X )  Y  kK  ( X k , Yk ) (5) 2
in which Sw is the common covariance matrix
For a model with T time steps, the conditional N mi
probability can be rewritten according to exponential SW   Σi (10)
i 1 m  N
feature functions F as The classification rule is given in eq. 11.
* *
1 T f ( x )  i : i  arg max d i ( x ) (11)
P (Y | X )  exp(  F ( y t , yt 1 , X )) (6) i
Z ( X ) t 1
The classification rule for LDA is very intuitive. This
Where F can be factorized to a weighted sum over a method requires a training phase, meaning the
computation of the discriminant functions and their
set of features  f n  parameters. The new data x can be classified simply by
solving the appropriate discriminant function for each
F ( y t , yt 1 , X )    n f n ( y t , yt 1 , X ). (7) class Ci and applying the classification rule (eq. 11).
n
IV. EXPERIMENTAL RESULTS AND DISCUSSIONS perfect measurement for the classifiers because the
In this section, we first give a description of the dataset contain dominant classes which appear more
dataset and provide details of our experimental setup and frequently than others. Therefore, the class accuracy
then we present and discuss the results. should be the primary way to evaluate an activity
classifier’s performance. However, in this paper we
A. Dataset report both the accuracy and the class accuracy.
We used Kastersen’s real world dataset [3]. This Sensors outputs are binary and represented in a
dataset were all recorded using a wireless sensor network feature space which is used by the model to recognize the
in homes with a single occupant in the house of a 26- activities performed. The “raw” sensor representation
year-old man. He lives alone in a three-room apartment gives a 1 when the sensor is firing and a 0 otherwise. We
where 14 digital state change sensors were installed. do not use the raw sensor data representation as
These sensors are attached to doors, cupboards, a observations, instead we use the “change point” and
refrigerator, and a toilet flush. The data collection lasted “last” representation which have been shown to give
for l=28 days. A total of 245 activity instances were much better results in activity recognition [3]. The
annotated by the participant with 2120 sensor events. A “change point” representation gives a 1 when the sensor
list of activities that were annotated with information reading changes. While the last sensor representation
about class distribution can be found in Table Ι. These continues to assign a 1 to the last sensor that changed
activities were chosen from the Katz ADL index [2]. state until a new sensor changes state.
Annotation was done by the subject himself at the same
time the sensor data was recorded. All unannotated time C. Results
slices are collected in a single activity annotated “Idle”.
We compared the performance of the C-SVM, CRF
and LDA on the imbalanced database in which minority
TABLE I. NUMBER (NB) OF ACTIVITIES, NUMBER (NB) OF class are all classes that take up 1% at most of percentage
OBSERVATIONS AND PERCENTAGE OF TIME ACTIVITIES OCCUR IN THE
HOUSE DATASET
of time. The others classes are the majority classes. These
algorithms are tested under MATLAB environment.
ADL Nb of Nb of Percentage SVM algorithm is tested with implementation LibSVM
activities observations of time (%)
[23].
Idle (Id) - 4627 11.5
Leaving (Le) 34 22617 56.4 Experiments were run using the “Changepoint+Last”
Toileting (To) 114 380 1.0 representation that is a concatenation of the (C.P) and
Showering(Sh) 23 265 0.7
(L.C) representation. The results obtained with this
Sleeping (Sl) 24 11601 29.0
Breakfast (Br) 20 109 0.3
representation in terms of accuracy and class accuracy for
Dinner (Di) 10 348 0.9 different methods are summarized in Table ΙΙ.
Drink (Dr) 20 59 0.2
TABLE II. ACCURACY AND CLASS ACCURACY FOR C-SVM, CRF,
B. Setup and Performance Measures AND LDA
We separate our data into a test and a training set
Methods Accuracy (%) Class(%)
using a “leave one day out cross validation” strategy [3].
C-SVM 93.4 75.5
In this strategy, a classifier is designed using (l - 1) days
CRF [3] 95.6 70.8
and evaluated on the one remaining day; this is repeated l
times, with different training sets of size (l - 1) and report LDA 93.5 72.4
the average performance measure. We get inferred labels
for the whole dataset by concatenating the results In our experiments, the C-SVM hyper-parameters (σ,
acquired for each day. C) have been optimized in the range (0.1-1) and (0.1-10)
respectively to maximize the class accuracy of leave-one-
As the activity instances were imbalanced between
day-out cross validation technique. Then, the best pair
classes, we evaluate the performance of our models with
parameters (σopt, Copt ) = (1, 1) are used. Using these
two types of measure similar to [3]. The accuracy was
values, we have found that the worst result by
measured by
Class=61.1%. In order to improve the results, we tried to
find the penalty parameters Copt (class) adapted for
in1[inf erred (i )true( i )] different classes with weighted Libsvm option [23] by
Accuracy  m (12)
maximising the class accuracy using cross validation, see
and the class accuracy was measured by table ΙΙΙ. We have found that the better result was then
Class=75.5%.
Class  1 
i 1 
N   n c inf erred ( i )  true ( i )
c c  (13)
N c 1
 mc 
TABLE III. PARAMETERS COPT ADAPTED FOR DIFFERENT CLASS
in which [a = b] is a binary indicator giving 1 when FOR C-SVM

true and 0 when false. m is the total number of samples, N ADL Id Le To Sh Sl Br Di Dr


is the number of classes and mc the total number of Copt 0.5 0.1 8 9 0.1 3 6 9
samples for class c.
Even though the accuracy is the typical way of We see that the minority class requires a large value
evaluating a classifier’s accuracy [3], it is not always the of Copt compared to the majority class. This fact induces a
classifier’s bias in order to give more importance to the and the three kitchen activities. Theses activities perform
minority ones. worst.
The table ΙΙ shows that C-SVM performs better in
terms of the class accuracy measure, while CRF performs TABLE VI. CONFUSION MATRIX FOR LDA. T HE VALUES ARE
better in terms of accuracy measure. C-SVM and LDA PERCENTAGES

have similar performances of accuracy. LDA outperforms Id Le To Sh Sl Br Di Dr


CRF for the class accuracy measure. Id 61.7 9.3 0.4 9.9 4.6 0.8 12.5 0.7
In order to find out which activities are relatively Le 0.8 98.4 0.1 0.3 0.0 0.1 0.1 0.0
harder to be recognized, we analyzed the confusion To 13.7 3.1 73.1 4.7 4.5 0.2 0.0 0.5
matrices of C-SVM, CRF and LDA methods in tables IV, Sh 7.1 0.0 3.4 89.4 0.0 0.0 0.0 0.0
V and VI respectively, which give information about the Sl 0.2 0.0 0.1 0.2 99.4 0.0 0.0 0.0
actual and predicted classification results given by the Br 24.7 0.0 1.0 0.0 0.9 52.3 7.3 13.7
classifiers. Di 36.0 1.7 0.3 0.0 0.3 16.3 38.5 6.9
Dr 17.0 3.4 0.0 0.0 0.0 3.4 10.1 66.1
The confusion matrix for C-SVM can be found in
table IV. The activities ’Leaving’, ’toileting’, D. Discussions
’Showering’, and ’Sleeping’ are better successfully
recognized comparatively with others activities. It can be In this section, we explain the difference in in terms
seen that the most confusion takes place in ’Idle’ activity, of performance between C-SVM, CRF and LDA. A
and the three kitchen activities ’Breakfast’, ’Dinner’ and multiclass C-SVM trains a several binary classifiers to
’Drink. Additionally, Theses activities perform worst and differentiate the classes according to the class labels and
are in general hard to recognize. optimise the parameter C for each class separately. As a
result, it is observed that this method achieves superior
results for minority class with respect to other methods.
TABLE IV. CONFUSION MATRIX FOR C-SVM. THE VALUES ARE The LDA may not work for non linearly separable
PERCENTAGES
dataset. It is found that the class accuracy of LDA is
Id Le To Sh Sl Br Di Dr inferior to C-SVM. The CRF model does not model each
Id 62.2 6.2 1.0 9.8 4.4 0.1 15.5 0.7 action class individually, but it models all the train
Le 0.83 98.0 0.6 0.3 0.0 0.0 0.2 0.0 samples as a whole. As a result classes that are more
To 10.2 2.0 83.1 2.6 1.0 0.5 0.2 0.2 dominantly present in the data have a bigger weight in the
Sh 7.5 0.0 4.5 88.0 0.0 0.0 0.0 0.0 CRF optimisation. This why CRF performs so much
Sl 0.0 0.0 0.8 0.1 99.0 0.0 0.0 0.0 better for the majority activities (’Idle’, ’Leaving’ and
Br 23.9 0.0 0.0 0.0 0.9 43.1 28.4 3.6
’Sleeping’), since theses activities take up 11% at least of
Di 23.0 0.3 0.5 0.0 0.0 4.3 68.1 3.7
the data. In other words, CRF overfits for these activities
since they occur more often in the dataset.
Dr 16.9 1.7 3.4 0.0 0.0 3.4 11.9 62.7
The recognition of the minority activities ’toileting’,
TABLE V. CONFUSION M ATRIX FOR CRF. T HE VALUES ARE ’Showering’, ’Breakfast’ ’Dinner’ and ’Drink’ is lower
PERCENTAGES
compared to ’Leaving’ and ’Sleeping’ activities. This is
Id Le To Sh Sl Br Di Dr mainly due to the fact that these activities are less
Id 80.4 12.3 0.4 0.4 4.3 0.0 2.0 0.2 represented in the training dataset (each take up 1% at
Le 1.1 98.4 0.2 0.1 0.1 0.0 0.1 0.0 most). However, the activities ’Idle’, and the three
To 7.6 12.4 69.5 1.8 8.7 0.0 0.0 0.0 kitchen activities gave the worst results compared to the
Sh 23.0 8.7 2.3 66.0 0.0 0.0 0.0 0.0
others activities. Most confusion occurs between the
Sl 0.1 0.0 0.2 0.0 99.7 0.0 0.0 0.0
’Idle’ activity and the three kitchen activities. In
particular, the activity ‘Idle’ often takes up a large
Br 32.1 0.0 0.9 0.0 0.9 55.0 9.2 1.8
amount of time slices but is usually not a very important
Di 33.0 7.2 0.6 0.0 0.6 4.3 53.4 0.9
activity to recognize. It might therefore be useful to
Dr 32.2 5.1 3.4 0.0 0.0 5.1 10.2 44.1
weigh this activity less. The kitchen activities are worst
for all methods because most of the instances of these
As we can see in table V, the activities ’Idle’, activities were performed in the same location (kitchen)
’Leaving’ and ’Sleeping’ give the highest accuracy for using the same set of sensors. In other words, it is
CRF method. The table show that CRF mainly performs observed that groups of similar activities are more
better for the ’Idle’ activity. Activities ’Breakfast’, separable if performed in different locations. For
’Dinner’ and ’Drink’ perform worse. Most confusion example, toileting and showering are more separable for
takes place in the ’Idle’ activity and the three kitchen all methods because they are in two different locations in
activities. Kasteren’s dataset [3]. Therefore the location of the
Finally, we report in table VI the confusion matrix for sensors is of great importance for the performance of the
LDA. The activities ’Leaving’, ’Showering’, and recognition system.
’Sleeping’ give the highest accuracy while the ’Toileting’
activity is less accurate than these activities. It can be V. CONCLUSION AND FUTURE WORK
seen that the most confusion takes place in ’Idle’ activity, In this paper, we showed the potential of
discriminative models named C-SVM, CRF and LDA for
activity recognition, by comparing their performance on [18] W. Lin, M.T Sun, R. Poovandran, and Z. Zhang. Human Activity
available real world dataset. Our experiments show that Recognition for Video Surveillance. International Symposium on
Circuits and Systems (ISCAS'08), Seattle, WA, May 2008.
CRF is more sensitive to overfitting on a dominant class
[19] A. Fleury, M. Vacher, N. Noury, “ SVM-Based Multi-Modal
than C-SVM and LDA. C-SVM outperforms CRF and Classification of Activities of Daily Living in Health Smart
LDA with respect to class accuracy. Using different Homes : Sensors, Algorithms and First Experimental Results,”
penalty parameters in the C-SVM formulation improves IEEE Transactions on Information Technology in
the low classification accuracy caused by imbalanced Biomedicine, Vol. 14(2), pp. 274-283, March 2010.
dataset. Therefore, the soft margin SVM for multi-class [20] S. W. Lee and K. Mase, “Activity and location recognition using
recognition is a feasible method. In our future works, it wearable sensors,” IEEE Pervasive Computing, Vol.1, No.3,
pp.24-32, 2002.
would be interesting to improve activities recognition
[21] J. Mercer, “Functions of positive and negative type and their
using hybrid discriminative – generative–models by
connection with the theory of integral equations,” Philosophical
combining C-Support Vector Machine (C-SVM)–Hidden Transactions of the Royal Society of London. Series A, vol. 209,
Markov Models (HMM) to capture the temporal pp. 415 – 446, 1909.
regularities and smoothness of activities. It has been [22] K., Veropoulos, C., Campbell and N., Cristianini. Controlling the
shown that hybrid models can provide better accuracy sensitivity of support vector machines. Proceedings of the
than their purely generative or purely discriminative International Joint Conference on AI, 55-60, 1999.
counterparts [24]. [23] C. C. Chang and C. J. Lin, LIBSVM: [Online]. Available:
http://www.csie.ntu.edu.tw/~cjlin-/libsvm/
[24] R. Raina, Y. Shen, AY. Ng, A. McCallum. Classification with
REFERENCES hybrid generative/discriminative models. In: S. Thrun, L. Saul, B.
[1] United Nations Population Division. World population prospects, Schölkopf, (eds), Advances in neural information processing
the 2002 revision, highlights, February 2003. systems 16. MIT Press, Cambridge, MA, 2004.
[2] S. Katz, T. D. Down, H. R. Cash, Progress in the development of
the index of ADL. Gerontologist, 10:20–30, 1970.
[3] T. van Kasteren, A. Noulas, G. Englebienne, and B. Krose.
“Accurate activity recognition in a home setting,” in UbiComp ’08
. New York, NY, USA: ACM, pp. 1-9, 2008.
[4] C. Bishop, Pattern Recognition and Machine Learning. Springer.
New York, ISBN: 978-0-387-31073-2, 2006.
[5] V.N. Vapnik. Statistical Learning Theory. New York: John Wiley
& Sons, 1998.
[6] A. Ng and M. Jordan. On discriminative vs. generative classifiers:
A comparaison of logistic regression and Naïve Bayes. In
Advances in Neural Information Processing Systems (NIPS), 2002.
[7] C. Sutton and A. McCallum. Introduction to Statistical Relational
Learning, chapter 1: An introduction to Condional Random Fields
for Relational Learning, page (Available online). MIT Press, 2006.
[8] R.A Fisher. The use of multiple measurements in taxonomic
problems. Annals of Eugenics, 7:179–188, 1936.
[9] E. M. Tapia, S. S. Intille, and K. Larson, “Activity recognition in
the home using simple and ubiquitous sensors,” in Proc. Pervasive
Computing, Vienna, Austria, pp. 158 – 175, April 18-23, 2004.
[10] B. Logan, J. Healey, M. Philipose, E. M. Tapia and S. S. Intille,
“A long-term evaluation of sensing modalities for activity
recognition,” In Ubicomp ’07, 483–500, 2007.
[11] W. Lao, J. Han, P. H. N. de With, “Automative video-based
human motion analysis for consumer surveillance system,” IEEE
Trans. Consumer Electronics, 55(2): 591-598, 2009.
[12] J. Lester, T. Choudhury, N. Kern, G. Borriello, B. Hannaford. “A
hybrid discriminative/generative approach for modeling human
activities,” In Proc. Int. Joint Conf. Artificial Intelligence,
Edinberg, UK, pp.766-772, 2005.
[13] M. Philipose, K. P. Fishkin, M. Perkowitz, D. J. Patterson, D.
Hahnel, D. Fox & H. Kautz, Inferring Activities from Interactions
with Objects, IEEE Pervasive Computing, Vol.3, No.4, pp. 50-57,
2004.
[14] S. Tsukamoto, H. Hoshino, and T. Tamura, “Study on indoor
activity monitoring by using electric field sensor,” in International
Conference of the International Society for Gerontechnology, Pisa,
Tuscany, Italy, June 4-7 2008.
[15] L. Lio, D. Fox, H. Kautz, “Extracting places and activities from
GPS traces using hierarchical conditional random fields,” Int. J.
Robotics Research, 26(1): 119-134, 2007.
[16] A. Schmidt, Ubiquitous Computing - Computing in Context. Ph.D.
Dissertation, Lancaster University, 2002.
[17] S. S. Intille, E. M. Tapia, J. Rondoni, and al., “Tools for studying
behavior and technology in natural settings,” In Ubicomp, pages
157–174, 2003.

You might also like