Daily living activity recognition based on statistical feature quality group selection

https://doi.org/10.1016/j.eswa.2012.01.164Get rights and content

Abstract

The benefits arising from proactive conduct and subject-specialized healthcare have driven e-health and e-monitoring into the forefront of research, in which the recognition of motion, postures and physical exercise is one of the main subjects. We propose here a multidisciplinary method for the recognition of physical activity with the emphasis on feature extraction and selection processes, which are considered to be the most critical stages in identifying the main unknown activity discriminant elements. Efficient feature selection processes are particularly necessary when dealing with huge training datasets in a multidimensional space, where conventional feature selection procedures based on wrapper methods or ‘branch and bound’ are highly expensive in computational terms. We propose an alternative filter method using a feature quality group ranking via a couple of two statistical criteria. Satisfactory results are achieved in both laboratory and semi-naturalistic activity living datasets for real problems using several classification models, thus proving that any body sensor location can be suitable to define a simple one-feature-based recognition system, with particularly remarkable accuracy and applicability in the case of the wrist.

Highlights

► New feature set selector based on discriminability and robustness criteria. ► Objective and qualitative idea about the discriminant power of every feature. ► Unprecedent powerful discriminant physical activity recognition features. ► Robust knowledge inference systems based on SVM and DT.

Introduction

The percentage of EU citizens aged 65 years or over is projected to increase from 17.1% in 2008 to 30.0% in 2060. In particular, the number of 65 years old is projected to rise from 84.6 million to 151.5 million, while the number of people aged 80 or over is projected to almost triple from 21.8 million to 61.4 million (EUROSTAT: New European Population projections 2008–2060). It has been calculated that the purely demographic effect of an ageing population will push up health-care spending by between 1% and 2% of the gross domestic product (GDP) of most member states. At first sight this may not appear to be very much when extended over several decades, but on average it would in fact amount to approximately a 25% increase in spending on health care, as a share of GDP, in the next 50 years (European Economy Commission, 2006). The effective incorporation of technology into health-care systems could therefore be decisive in helping to decrease overall public spending on health. One of these emerging health-care systems is daily living physical activity recognition.

Daily living physical activity recognition is currently being applied in chronic disease management (Amft and Tröster, 2008, Zwartjes et al., 2010), rehabilitation systems (Sazonov, Fulk, Sazonova, & Schuckers, 2009) and disease prevention (Sazonov et al., 2011, Warren et al., 2010), as well as being a personal indicator to health status (Arcelus et al., 2009). One of the principal subjects of the health-related applications being mooted is the monitoring of the elderly. For example, falls represent one of the major risks and obstacles to old people’s independence (Najafi et al., 2002, Yu, 2008). This risk is increased when some kind of degenerative disease affects them. Most Alzheimer’s patients, for example, spend a long time every day either sitting or lying down since they would otherwise need continuous vigilance and attention to avoid a fall.

The registration of daily events, an important task in anticipating and/or detecting anomalous behavior patterns and a primary step towards carrying out proactive management and personalized treatment, is normally poorly accomplished by patients’ families, healthcare units or auxiliary assistants because of limitations in time and resources. Automatic activity-recognition systems could allow us to conduct a completely detailed monitoring and assessment of the individual, thus significantly reducing current human supervision requirements.

The primary difficulty in activity recognition lies in designing a system the reliability of which is independent of the person carrying out the exercise or the particular style of execution of the activity in question. Complexity is further increased by distortion elements related to system monitoring and processing, along with the random character of the execution. Most studies to date have been based on laboratory data (i.e., involving direct supervision by the researcher) and have achieved successful recognition of the most prevalent everyday activities (lying, sitting, standing and walking: Aminian et al., 1999, Karantonis et al., 2006, Maurer et al., 2006, Ravi et al., 2005). Nonetheless, the apparently good recognition results obtained during supervised experiences cannot be extrapolated to habitual real-life conditions (Könönen, Mäntyjärvi, Similä, Pärkkä, & Ermes, 2010).

The ideal scenario would be a naturalistic monitoring context consisting of a scenario with no intervention on the researcher’s part and without the subject’s cognitive knowledge about the exercise conducted, but unfortunately this is currently unfeasible. Some studies have applied a so-called semi-naturalistic approach (Bao and Intille, 2004, Ermes et al., 2008, Foerster et al., 1999, Pirttikangas et al., 2006, Uiterwaal et al., 1998), an intermediate between laboratory and naturalistic monitoring based on the inference of the hidden activity through the proposal of a related exercise, thus minimizing the subject’s awareness of the true nature of the data being collected. This approximation is somewhat more realistic than laboratory experimental setups.

The classic method for activity identification is based on three main stages: feature extraction (e.g., statistical features (Baek et al., 2004, Maurer et al., 2006, Ravi et al., 2005), wavelet coefficients (Nyan, Tay, Seah, & Sitoh, 2006; Preece et al., 2009, Preece et al., 2009) or other custom-defined coefficients (He et al., 2008, Mathie et al., 2003)), feature selection (e.g., principal or independent component analysis (Mantyjarvi, Himberg, & Seppanen, 2001), forward–backward selection (Pirttikangas et al., 2006), correlation (Maurer et al., 2006), etc.) and classification (primarily supervised learning approaches such as artificial neural networks (Engin et al., 2007, Parkka et al., 2006, Zhang et al., 2005), support vector machines (Begg and Kamruzzaman, 2005, Parera et al., 2009, Sazonov et al., 2009), Bayesian classifiers (Bao and Intille, 2004, Wu et al., 2007) and hidden Markov models (Minnen et al., 2006, Sazonov et al., 2011), among others). For a detailed review of classification techniques used in activity recognition the reader is referred to Preece et al., 2009, Preece et al., 2009.

Evidently, all these stages are important, but in this work we want to emphasize the importance of selecting the most interesting features to improve the efficiency of the subsequent pattern recognition systems, especially bearing in mind the rather discouraging results obtained with semi-naturalistic data. It is well known that a large number of features are directly translated into numerous classifier parameters, so keeping the number of features as small as possible is in line with our desire to design classifiers with good generalization capabilities, the best scenario being a knowledge inference system defined by just a few features. Consequently, we propose here an automatic method to extract a subset of the most important features to be used in activity recognition, which is especially suitable for looking for optimum single-feature classifiers with multiclass absolute discrimination capability.

The rest of the paper is organized as follows: Section 2 contains a description of the experimental setup, preprocessing process, features extracted from the data and the proposed rank-based feature selection method. Section 3 presents the results obtained, including a comparison of the performance of several different approaches. These results are subsequently discussed in Section 4 and our final conclusions are summarized in Section 5.

Section snippets

Experimental setup

Our experimental setup starts from a set of signals corresponding to acceleration values measured by a group of sensors (accelerometers) attached to different strategic parts of the body (hip, wrist, arm, ankle and thigh) for several daily activities1 following both laboratory and semi-naturalistic monitoring schemes (Bao & Intille, 2004). Our study is focused on the four most common

Results

Having described the method we go on in this section to present the results corresponding to the particular activity-recognition problem analyzed.

Features ranked for laboratory data by applying RQG-DR for oth = 0 are shown in Table 2 (an equivalent table was obtained for semi-naturalistic data). For example, the fifth central statistical moment, the maximum and the range of the energy spectral density and the spectrum amplitude were ranked in the quality group #11. These features were able to

Discussion

The aim of this study has been to assess the feasibility of activity recognition systems based on a minimum defining feature set, taking into account a huge set of possible discriminant features, many of them disregarded in previous works. An extensive initial feature set is analyzed using several feature selection methods.

Knowledge inference systems designed for both laboratory and semi-naturalistic monitoring contexts achieve quite good results using several methods, as shown in Fig. 5a–c.

Conclusions

We have described a direct application of feature selection methods applied to daily physical activity recognition systems. An efficient classification method requires a productive and limited feature set, thus requiring an efficient selection process since the initial set of possible candidates is huge. We have designed a highly accurate feature selector based on statistical discrimination and robustness criteria, with very low computational and resource requirements, which represents a

Acknowledgements

This work was supported in part by the Spanish CICYT Project TIN2007-60587, Junta de Andalucia Projects P07-TIC-02768 and P07-TIC-02906, the CENIT project AmIVital, of the “Centro para el Desarrollo Tecnológico Industrial” (CDTI- Spain) and the FPU Spanish grant AP2009-2244.

We are grateful to Prof. Stephen S. Intille, Technology Director of the House_n Consortium with the MIT Department of Architecture for the experimental data provided.

References (46)

  • L. Bao et al.

    Activity recognition from user-annotated acceleration data

    Pervasive Computing

    (2004)
  • C.V.C. Bouten et al.

    A triaxial accelerometer and portable data processing unit for the assessment of daily physical activity

    IEEE Transactions on Biomedical Engineering

    (1997)
  • N. Cristianini et al.

    An introduction to support vector machines and other kernel-based learning methods

    (2000)
  • R.O. Duda et al.

    Pattern classification

    (2001)
  • M. Ermes et al.

    Detection of daily activities and sports with wearable sensors in controlled and uncontrolled conditions

    IEEE Transactions on Information Technology in Biomedicine

    (2008)
  • European Economy Commission Directorate-General for Economic and Financial Affairs (2006). The impact of ageing on...
  • J. Fahrenberg et al.

    Assessment of posture and motion by multichannel piezoresistive accelerometer recordings

    Psychophysiology

    (1997)
  • He, Z., Liu, Z., Jin, L., Zhen, L.-X., & Huang, J.-C. (2008). Weightlessness feature – a novel feature for single...
  • Heinz, E. A., Kunze, K.-Steven, Sulistyo, S., Junker, H., Lukowicz, P., & Tröster, G. (2003). Experimental evaluation...
  • D.M. Karantonis et al.

    Implementation of a real-time human movement classifier using a triaxial accelerometer for ambulatory monitoring

    IEEE Transactions on Information Technology in Biomedicine

    (2006)
  • Kern, N., Schiele, B., & Schmidt, A. (2003). Multi-sensor activity context detection for wearable computing. In...
  • Laerhoven, K. V., & Gellersen, H.-W. (2004). Spine versus porcupine: A study in distributed wearable activity...
  • S.-W. Lee et al.

    Activity and location recognition using wearable sensors

    IEEE Pervasive Computing

    (2002)
  • Cited by (105)

    • A feature extraction model based on discriminative graph signals

      2020, Expert Systems with Applications
      Citation Excerpt :

      As one of the most classical subjects of artificial intelligence and expert systems, pattern recognition techniques find wide applications in many fields, such as medical studies (Mahsa, Seyed, & Akhavan, 2019; Zheng, Sang, & Lam, 2014), detection (Banos, Damas, Pomares, Prieto, & Rojas, 2012; Li et al., 2011), recommender system (Phuong, Lien, & Phuong, 2019; Portugal, Alencar, & Cowan, 2018), classification (Brown & Mues, 2012; Mironczuk & Protasiewicz, 2018), etc.

    View all citing articles on Scopus
    View full text