Online classifier construction algorithm for human activity detection using a tri-axial accelerometer

Abstract

This paper presents an online construction algorithm for constructing fuzzy basis function (FBF) classifiers that are capable of recognizing different types of human daily activities using a tri-axial accelerometer. The activity recognition is based on the acceleration data collected from a wireless tri-axial accelerometer module mounted on users’ dominant wrists. Our objective is to enable users to: (1) online add new training samples to the existing classes for increasing the recognition accuracy, (2) online add additional classes to be recognized, and (3) online delete an existing class. For this objective we proposed a dynamic linear discriminant analysis (LDA) which can dynamically update the scatter matrices for online constructing FBF classifiers without storing all the training samples in memory. Our experimental results have successfully validated the integration of the FBF classifier with the proposed dynamic LDA can reduce computational burden and achieve satisfactory recognition accuracy.

Introduction

Human-computer interaction (HCI) is a notable discipline that bridges the gap between users and computer systems, and has increasingly being recognized as an indispensable component of daily life. One of the key techniques in HCI is pattern recognition since users’ intentions can be recognized by recognition techniques without using the traditional input devices of computer systems. Among various pattern recognition issues, activity detection/recognition is an emerging technique which can recognize human activities or gestures via computer systems. Signals for recognition can be obtained from different kinds of sensors or detectors such as electromyography (EMG), audio sensors, image sensors, and accelerometers. Due to the rapid development of sensor technology and the omnipresence of reasonable low-cost high-performance personal computers, research studies on human activity detection/recognition have been grown up rapidly in diverse fields including biomedical engineering, medical nursing, and interactive entertainment [1], [2], [3], [4]. Among the aforementioned sensors for activity recognition, accelerometers can return a real-time measurement of acceleration along the x-, y- or z-axis for human activity detection. Due to advanced miniaturization techniques, accelerometers can be embedded within a wearable device and the generated data can be transmitted wirelessly to a mobile computing device. This greatly reduces users’ awareness and possible discomfort during the process of data collection and recognition. From the report of the world inertial sensor’s market [5], accelerometers have been applied to many commercial applications such as toys and pedometers, auto key stoning for projection systems, video game control, and human machine interface (HMI) in consumer electronic devices. To name a few, Apple’s iPhone integrated with accelerometers can automatically adjust the display direction of its monitor. Samsung proposed a novel gestured-based interaction using tri-axial accelerometers and realized it into a commercialized mobile phone (SCH-S310). Users can command the phone by waving the phone in the shape of digits and alphanumeric symbols. Nintendo’s Wii embedded tri-axial accelerometers in its controllers to detect motions which can provide an innovative game play experience. The runaway success of Wii confirms the feasibility and commercial potential of activity recognition using accelerometers. Therefore, in this paper we focused on the development of an effective construction algorithm for activity detection using accelerometers.

Traditionally, human activity recognition has been performed mainly by statistical methods [6]. HMMs have been regarded as a statistical modeling tool that can be applied to modeling time-series with spatial and temporal variability. Jamie et al. [7] used both body-worn microphones and accelerometers to recognize continuous human activities (sawing, hammering, filing, drilling, grinding, sanding, opening a drawer, tightening a vise, and turning a screwdriver). They used linear discriminate analysis (LDA) and HMMs for activity classification based on the detected segments from sound channels and accelerometers, respectively. Kallio et al. [8] utilized interpolation and vector quantization to generate a set of features based on discrete time series. Sixteen ergodic HMMs were adopted to recognize sixteen different human gestures. Another alternative approaches are decision-tree-based classification methods which have been successfully used for many activity recognition problems [9], [10], [11], [12]. To name a few, Bao et al. [9] introduced four recognition methods, including the decision table, instance-based learning (IBL), C4.5 decision tree and naı¨ve Bayes classifiers, to recognize 20 human daily activities (walking, watching TV, running, reading, bicycling, etc.) The results of activity recognition were based on acceleration data from five biaxial accelerometers placed on 20 subjects in five positions (arm, wrist, hip, angle and thigh), and the data were collected under a semi-naturalistic condition. The C4.5 decision tree obtained the best performance with an overall recognition accuracy rate at 84%. The decision tree method usually separates static activities from dynamic activities first, and a more detailed subclassification is made at the following hierarchical structures [10], [11], [12]. The advantages of decision trees are that they are simple, apparent, and fast in reasoning. However, traditional decision trees need to keep a large amount of data in each of their non-terminal roots, which increases the required memory space and computation time.

Among the aforementioned studies, feature dimension reduction which is an important procedure in pattern recognition was not considered. In a recognition problem, a high dimensional feature set may result in two problems: (1) Some features are irrelevant or redundant and do not provide supportive information to significantly improve recognition accuracy; and (2) the training of classifiers in a high dimensional space is difficult and requires more computational time. Therefore, before performing the recognition task, it is desirable to reduce the dimension of the feature set, to find out important features which are helpful to recognition, and to retain the characteristics of the original data distribution in the space spanned by the important features. Principal component analysis (PCA) is one of the well-known dimension reduction methods. The basic idea of the PCA is to seek a projection that transforms the original features into a lower dimensional space and preserves most information of the original data. Although the PCA is good at seeking the best representative data projection, it may not be useful for data discrimination. Unlike the PCA, the LDA performs well in seeking a suitable projection for data discrimination. That is, the LDA aims to find out an effective transformation for separating the data distribution into different classes while minimizing the data distribution of the same class in a new space. Although the effectiveness of the LDA in recognition or classification problems has been validated in many studies [13], [14], the feasibility of the LDA for the applications at which complete training data sets are not available beforehand still remains some challenges. To enable the LDA with dynamic data processing capability, several researchers [13], [14], [15], [16] have devoted their efforts to modify LDA for incrementally updating the discriminant components when more data becomes available. To name a few, Kim et al. [15] presented a new incremental learning method which employed sufficient spanning set approximation technique in each update step and made the computation of eigenproblem more efficient. An increment LDA (ILDA) learning algorithm which combines reconstructive and discriminative information was proposed in [14]. The proposed method allows incrementally updating the current presentation of the existing classes with new instances as well as adding new classes to the existing classes. Pang et al. [16] derived an intuitive method to update the scatter matrices of the LDA for large data streams. Similar to the idea in [16], we proposed a dynamic LDA which is capable of updating the scatter matrices continuously without saving all the training data points. Our proposed method utilizes only the statistical information of class distributions and can be operated in both “incremental” mode and “decremental” mode as well. That is, our method is capable of updating the existing classes by adding additional new data as well as creating new classes (or deleting some classes) to (from) the existing classes. The proposed dynamic LDA can compute scatter matrices as accurately as the original LDA operated in a batch mode. In our design of classifiers, we utilized the information obtained from the dynamic LDA to construct a fuzzy basis function (FBF) classifier for activity recognition.

The organization of this paper is as follows. Section 2 introduces the conventional LDA formulation and presents the derivation of the proposed dynamic LDA. The detailed information about the proposed online classifier construction is presented in Section 3. Section 4 provides the experimental results that validate the effectiveness of the proposed strategy and classifier. Conclusions are given in the last section.