Luận văn apply incremental learning for daily activity recognition using low level sensor data

‘VIRTNAM NATIONAL UNIVERSITY, HANOT UNIVERSITY OF ENGINEERING AND TECIINOLOGY TA VIET CLONG APPLY INCREMENTAL LEARNING FOR DAILY ACTIVITY RECOGNITION USING LOW LEVEL SENSOR DATA M

‘VIRTNAM NATIONAL UNIVERSITY, HANOT UNIVERSITY OF ENGINEERING AND TECIINOLOGY

TA VIET CLONG

APPLY INCREMENTAL LEARNING FOR DAILY

ACTIVITY RECOGNITION USING LOW LEVEL

SENSOR DATA

MASTER THESIS OF INFORMATION TECHNOLOGY

Ha Noi, 2012

‘VIRTNAM NATIONAL UNIVERSITY, HANOT UNIVERSITY OF ENGINEERING AND TECIINOLOGY

TA VIET CLONG

APPLY INCREMENTAL LEARNING FOR DAILY

ACTIVITY RECOGNITION USING LOW LEVEL

SENSOR DATA

Major: Computer Science Code: 60.48.01

MASTER THESIS OF INFORMATION TECHNOLOGY

Supervised by Assoc, Prof Bui The Duy

Ha Noi, 2012

Table of Contents

a

Introduction

11 Overview

12 Our Works

13 Thesis Outline

Itclated Work

21 Daily Activity Recognition im Smart Tome System

22 Dally Activity Recoguition approach

2.3 Incremental learning model

Framework for Activity Recugnition in the Home Environment

3.2 Date annotation ca TH ng ng ng và và 3.3 Feature extraction

3.4 Segmentation

Growing Neural Gas model

4.1 GNG structure - -

4.11 Competitive Ilebbian Learning 41.2 Weights Adapting

4.2 Using GNG network for supervised learning

4121 Separated Weighted Tuelid - 4.2.2 Reduce Overlapping Regions by Using a Loval Exror Threshold

ew Node Jusertion

Radial Basic Function nctwork

5.1 Standard Radial Basic Mimerion

3.2 Incremental Radial Basic Function "

‘TABLE OF CONTENTS

6 Experiment and Result

7 Couclusion

7.1 Summary

73 Tanture works,

31

35

35 36

Apply Incremental Learning for Daily Activity Recognition Using Low Level Sensor Data

Abstract

Daily activity recognition is an important task of many applica- tions, especially in an environment like smart home ‘I'he system needs

to be equipped with Uke recoguition ability se Uhat it cau automat

ically adapt to resident's preferences So far a great deal nmmber

of learning models has been proposed to classify activities Llowever,

znost of Went uly work well iu off-line inanuer when training data is

known in advance It is known that peaple’s living habits change aver

Lime, therefore a learning beeimique Uhat should learn uew knowledge

when there is new data in great demand

In this thesis, we Improve an existing incremental leaning madel

to solve this probleu The model, whieh is traditionally uscd in uusu-

pervised learning, is extended for classification problems Tneremental

learning strategy by evaluating the error classification is applied when

deciding whether a new class is generated, A separated weighted Eu

clid is used ta measnre the distance between samples hecanne the lai

variauce of injorimalion contains in [eeture vectors To avoid coustaut

allocation of new noctes in the overlapping region, an arlaptive inser

tion constraint is added Finally, an experiment, is carried tia assess

ius performance The resulls show Lal the proposed anethud is better

than the previous one The proposed method can he integrated into

a smart system, which then pro-actively adapts itself to the changee

of human daily activity pattern

Smort home is considered as a part of Ubiquitous Computing trond It aims

to eurieh the home cnvironmcul with intelligeut deviews with Lhe purpose of

supporting the residents living inside the environment '4), Understanding the users’ activities is one of the key fearures of a smart home system Regular

activities which are performed in the home can he divided inte two types One type of activities is characterized as simple gestures of some part of

the body such as running or walking I'he other one is the set of actions which become a pattern in users’ claily routines Some examples of this type

are reading, cooking or watching television Recognizing these activities will

provide useful information [or understanding the environment context, In our thesis, we focus on Lhe problem of recognizing Une second types of the

activities

The most common data sonree ìn the smart hame sys

sensor information 4] 'here have been many propose approaches for recog-

nizing activities from low level sensory data (2U, 3, 21|) One of the proper-

tics of daily activitics in the home environment is its variation through time bevwuse Une uscr'y habits are change in real life, To deal with Uhis prob-

Ten, previous spprouch must have bo brain the models again when there are changes in activities patierus However, vis is a resource consuming process,

Tn onr thesis, we apply incremental learning model into the problem of activity recognition to resaive the ahove problem The incremental learning

models are based on the Growing Neural Gas (GNG) network [8] We extend

the growing neural gas model for activity recognition in two ways ‘I'he first approach is to use multiple GNG networks Mor each activity class, we train

a separated network using samples of the class Technically, the weak point

of the traditional growing neural is its network size constantly grows based

on the error evaluation for new sample insertion This learning, strategy encounters the performance problem in case classes averiap We propose

the use of another metric for measure distance nm the GNG network and a

constraint to the growing condition with the purpose ta make the models more balance between each class ‘I'he second approach is to create a Radial

Basie Function (RBF) network [16] from the GNG network This approach

iy similar Lu the method proposed in [7]

We carry oul an experiment to compare both models Performance is evaluared on real daily activity dataset The experimental results show that the task of recognizing daily activity can achieve good result with incremental learning approach

‘The remaining of thesis is organized as follows In Section 2, we present the works related to activity recognition and incremental Icarning In See- tion 3, we describe more detuil of activily revoguition problems including

m is low-level

2

date acquisition, data scgmentation and feature extraction In Section 4, we introduce the structures of the GNG network and the way of using, it far daily activity recognition problem Tn Section 5, the incremental version of the RIM network based on the GNG network is presented An experiment: with real data set is presented in Section 6 In Section 7, we discuss the summary of our thesis and future works

In recent yours, there are many rescurches in building a system for smart

liome environment In these systems, the aclivily revoguition iy considered

asa part of the context recognition module While the context recognition module is refer to understand a wide range of knowledge in the intelligent environment, the main purpose of activity recognition is to extract the in formation of what the resident is doing ‘he task of recognizing activities

depends on which sensory data the system can perceive from the real world

There are w variety of sensur (ypes in a smart home system Sensory ean be

used for pusilion measurement, delecting motion detecting sound, reading

temperature ‘These sensors usnally create many data streams 14], which

have to be combined in order to produce more useful information Compar-

ing to other types of reoivers like cameras or microphones, using low level sensors offers low cost im building the sensor network and transparency ‘lhe data generated by low level sensors is quite easy ta colleet anc process in comparison to other types of device like camera

2.1 Daily Activity Recognition approach

There are nnany proposed models for recoguizing daily activities In [20], they built an wetivity rocognivion model based on Nuive Bayes clusiliers Lu [23]

a multilayer model is used to detect walking movements from passive infrared motian detectors placed in the environment Tn [3], they proposed a frame- work for activity recognition including feature extraction, feature selection and predicting models The framework use the four popular machine learn- ing methods including Laves Belief Networks, Artificial Neural Networks Sequential Minimal Optimization and Boosting

One of the propertics of daily activity is the change of their patterns becanse of the vary of users’ habits This makes the derision boundary ean

change through time Proposed methods requires to troin the models sgain when the changes ocent Incremental learning approach can be nsed to solve this problem

2.2 Incremental learning model

‘The approach of incremental learning is aim to handle the problem of learning new knowledge while maintaining the previous one [11 In unsupervised approach, the well known clustering algorithm k-means [13 can be learned on-line by a weight update rule An artificial neural network, namely Self- Organized Map (SOM) (121, is presented as a unsupervised learning method

using incremental approach More flexible networks with similar approach ure Growing Cell Structures ‘7 aud Growing Neural Cas [8]

In supervised leurning, there ure several efforts to adapt the exist ofl Tine method to use im incremental training With the inspiring of adap-

tive host algorithm{AdaPioost) [6], [L7] proposed an incremental learning

method, namely Learn++ Tt uses the Multilayer Pereeptron network aa a week learner to generate a number of weak hypotheses Radial Hasic Func

tion (RBF) networks [16] combine a local representation of the input space

aud a pereeptron layer tu fit the input puttcra with its target [8] proposed a

method lo insert new nodes inte the RBF network, Using this upproudh, the RBF network is capable for learning overtime Another approach is Fuzzy ARTMAP |2| which based on Adaptive Resonance Theory (ART) networks

Home Environment

In this section, we present the framework of daily activity recognition using low level sensor data in the home envizonment Its mein purpose is to map the datu from the cuviromment to a sequence of uetivities We adapt the proposed fumework in [3] (o emphasize Une incremental icurning propertios

of the framework Figure 1 illustrates the steps in an activity recognition module using tow-level sensor data ‘I'he information of the surrounding en- vironment comes under a stream of censor events ‘hen, the stream can be annotated and split into lebelled samples ‘I'he labeied samples are then pro: vided to the inercmental learning model for training, In online recognition the stream is segmented into separated segucuces of scuser events Each

4

Labelled samples

split

data

Classifier

Sequence of activities

Figure 1: Framework for activity recognition

separated sequenee is then classified by the learned model By applying in- cremental learning, the model is trained continnonsly aver the Tife span of the system whenever there is new labelled data

3.1 Duta Acquisition

Tn the data acquisition phase, sensors are implemented around the home

The sensors are low level, which are differ from cameras or microphones

‘They continuously monitor the environment for some specific information

There ure many lypes of information in the howe environment that ew be monitored by slate change sensor such as door, pickiug object or lemperalure

Choosing the types of sensor are ineInded totally depends on the system's design

The data, from each low level sensor is collected and combined into a stream of data at a central server Mach sensor's signal is considered as an

event in the stream An example of event stream 5] is given in table | Hach

event has a time stamp, the name of the sensor and its vafue ‘here are

several platforms for collecting and processing the data stream in the smart home system such ag Motes plalform “I] or O8CT plallorit [10

on

‘Table 1: An example of sensor events in data stream

Date Time — Sensor | Valne |

2009-10-16 | 08:50:00 M026 | OX 2009-10-16 | 08:50:01 M026 | OFF |

2009-10-16 | 08:50:02 M02’ | ON

2009-10-16 | 08:60:13 M026 | OFF 2009-10-16 | 08:50:17 M026 | OFF |

For the purpose of training model, the stream data needs to he segmented

and labelled into separated activities In a smart home system, data can be annotated dircetly by using deviccs such as Personal Digital Assistant [20]

An allernative way iy Iubelling the slream data indi through » post processing step The users can label their aclivilies with a visualization ool

[3| Because the data is captured from various sensors and span aver a long

time, it is difficult to determine the right starting and ending point, of an

activity After this step, each sensor event in the stream dara is added with

an optional action label

Trecanse the daily activities in the home environment nsually happen around

a specific area, the motion sensors can produce good features for differing the uetivily classes In [3], the activated motion sensors during the period

of an activity iy included inlo the Íewbure voctor The [ewbure veubor also uses high contextual information such as day of week, lime of day, previews ackivitiy, next activity and energy consumption This approach achieves a high accnracy with learning models anch as multilayer perceptron or decision

tree

In our approach, we use a similar method to extract feature for training with online learning model A feature vector of an activity sample ineludes:

®@ Length of the activity in second

* Part of the day when the activity happens, It is divided into six sepa- rated paris: morning, noon, afternoun, ceening, night, lutenight,