I-vector analysis for Gait-based Person Identification using smartphone inertial signals

https://doi.org/10.1016/j.pmcj.2016.09.007Get rights and content

Highlights

  • Gait Recognition using I-vector transformation.

  • Analysis of different session compensation techniques: LDA and PLDA.

  • Comparison with using Gaussian Mixture Model-Universal Background Model.

Abstract

This paper describes and evaluates an i-vector based approach for Gait-based Person Identification (GPI) using inertial signals from a smartphone. This approach includes two variability compensation strategies (Linear Discrimination Analysis (LDA) and Probabilistic LDA) for dealing with the gait variability due to different recording sessions or different activities carried out by the user. This study uses a public available dataset that includes recordings from 30 users performing three different activities: walking, walking-upstairs and walking-downstairs.

The i-vector approach is compared to a Gaussian Mixture Model-Universal Background Model (GMM-UBM) system, providing significant performance improvements when incorporating the PLDA compensation strategy: the best result reports a User Recognition Error Rate (URER) of 17.7%, an Equal Error Rate (EER) of 6.1% and an F1-score of 82.7% with 30 enrolled users. For less than six enrolled users, the URER error decreases to 5%.

Introduction

The research in multi-sensor networks has increased significantly due to the sensor price reduction. These networks typically include cameras, Indoor Location Systems (ILS), microphones, etc. Thanks to the information obtained from the sensors, computer based systems can make more intelligent actions adapting their behavior to contextual conditions. Recognizing different situations can be seen as a classification problem where some features are obtained from sensor signals and several mathematical models are generated for every different situation.

Thanks to the increment of sensor networks, the number of possible research areas has also increased rapidly. One of these areas is Gait-based Person Identification (GPI): recognition of a user identity based on his/her gait. This area of research has received a lot of attention in the last years due to the high number of promising applications (people monitoring and supervision).

Initial studies (since 1967) described gait as a personal characteristic. In one of these experiments, Johansson  [1] attached moving lights onto human subjects on all the major body parts. The observers could recognize the biological patterns of gait from the Moving Light Displays (MLDs), even when some of the markers were removed. These studies indicated that gait could be a candidate for biometric recognition.

Nowadays, gait recognition is considered one of the most active areas of biometrics  [2]. Gait offers significant advantages compared to other biometric traits: unobtrusiveness (physical contact with the subject is not required as in the case of fingerprint acquisition), user-friendliness (the gait of a person can be captured unobtrusively and continuously, unlike fingerprinting or retina scans), and security (the gait of an individual is difficult to mimic (see  [3])). Other biometric modalities such as fingerprints are relatively easy to copy and the security depends much more on the resistance of the sensor against fake inputs (see  [4]). But gait has the problem that it can vary depending on the session or the activity carried out by the user (walking, walking-upstairs or walking-downstairs, for example). In order to face this problem, it is necessary to develop robust pattern recognition algorithms that include variability compensation strategies. This aspect is addressed in this paper and it is an important contribution compared with previous works.

This paper proposes and evaluates an i-vector based approach for GPI using inertial signals from a smartphone. This work has been done using a public dataset  [5]. The main contributions of the paper are the followings:

  • The first contribution is a detailed description of an i-vector based approach adapted to inertial signal processing. These signals are recorded from an accelerometer and a gyroscope embedded in a smartphone.

  • Secondly, the paper includes an analysis of the different algorithm parameters in order to report the best system configuration.

  • The third contribution consists of a comparison between the i-vector approach and a Gaussian Mixture Model-Universal Background Model (GMM-UBM) approach described in  [6].

  • Finally, this work proposes two variability compensation strategies for dealing with session and activity variability: Linear Discriminant Analysis (LDA) and Probabilistic Linear Discriminant Analysis (PLDA). Thanks to these compensation strategies, it has been possible to obtain significant performance improvements. As far as authors know, this study shows the best results over this dataset on GPI.

The paper is organized as follows. Section  2 presents the state of the art. Section  3 describes the GPI approaches analyzed in this study. Section  4 details the experiments carried out in this work including the database description and the evaluation metrics. Section  5 discusses the system applicability and, finally, Section  6 summarizes the main conclusions of this work.

Section snippets

Background

Gait-based Person Identification can be performed by using the information obtained from different types of sensors: video camera, Microsoft Kinect output, accelerometers, etc. Video cameras have been one of the most extended sensors for gait recognition  [7]. In computer vision, the general framework for automatic gait recognition consists of several steps: subject detection, silhouette extraction, feature extraction, feature selection and person identification. Once moving subjects are

Gait-based Person Identification

This section describes the two methods considered in this study: a GMM-UBM approach and an i-vector approach. This work has been performed using the Microsoft Research Identity Toolbox  [40]. This toolbox contains a collection of MATLAB tools and routines that can be used for research and development in speaker recognition. These tools have been adapted to process inertial signals. In this work, several moving activities have been considered for GPI: walking, walking-upstairs and

Experiments

This section describes all the experiments carried out in this work. Sections  4.1 Data base description and experiments configuration, 4.2 Evaluation metrics used in the experiments describe the database, the experiment configuration and the evaluation metrics used in the experiments. Section  4.3 shows the experiments (on the validations subsets) for tuning the main parameters of the systems. Finally, Section  4.4 includes the final results (on the test subsets) and a comparison with other

System applicability

This section discusses the applicability of the GPI approach described in this work. In general, GPI approaches require the user to perform a walking activity during several seconds, so they cannot be used in all cases: for example, with people on a wheelchair. The need of performing a walking activity makes difficult the use of GPI approaches in a one-step authentication process, like logging into an account, for example. In the other hand, GPI approaches can be used as a continuous background

Conclusions

This paper has proposed a new approach for Gait-based Person Identification using inertial signals from a smartphone: i-vector analysis with PLDA compensation. This approach consists of six main steps: feature extraction, GMM-UBM training, T (total variability) matrix computation, i-vector extraction using the T matrix, PLDA variability compensation, and finally, the scoring module that compares the current i-vector with all the i-vectors from the previously enrolled users. This work has

Acknowledgments

The work leading to these results has been supported by ASLP-MULÁN (TIN2014-54288-C4-1-R) and NAVEGABLE (MICINN, DPI2014-53525-C3-2-R) projects.

The authors want to thank the UCI Machine Language Repository and especially the researchers who kept the records and developed the Human Activity Recognition Using Smartphones Dataset: researchers from Smartlab (Non-Linear Complex Systems Laboratory, DITEN, Università degli Studi di Genova) and Technical Research Centre for Dependency Care and

References (47)

  • T. Fawcett

    An introduction to ROC analysis

    Pattern Recognit. Lett.

    (2006)
  • G. Johansson

    Visual perception of biological motion and a model for its analysis

    Attention Percept. Psychophys.

    (1973)
  • X. Qinghan

    Technology review - Biometrics-Technology, Application, Challenge, and Computational Intelligence Solutions

    IEEE Comput. Intell. Mag.

    (2007)
  • B.B. Mjaaland, P. Bours, D. Gligoroski, Walk the walk: attacking gait biometrics by imitation. in: Proceedings of the...
  • T. Matsumoto, H. Matsumoto, K. Yamada, S. Hoshino, Impact of artificial ”gummy” fingers on fingerprint systems. in:...
  • D. Anguita, A. Ghio, L. Oneto, X. Parra, J.L. Reyes-Ortiz, A public domain dataset for human activity recognition using...
  • T.K. Lee et al.

    A comprehensive review of past and present vision-based techniques for gait recognition

    Multimedia Tools Appl.

    (2014)
  • S.X. Yang et al.

    Variability and similarity of gait as evaluated by joint angles: implications for forensic gait analysis

    J. For. Sci.

    (2014)
  • G. Ariyanto, M.S. Nixon, Marionette mass–spring model for 3D gait biometrics. in: 5th IAPR International Conference on...
  • J. Shotton, T. Sharp, A. Kipman, A. Fitzgibbon, M. Finocchio, A. Blake, M. Cook, R. Moore, Real-time human pose...
  • M. Milovanovic et al.

    Walking in colors: human gait recognition using kinect and CBIR

    IEEE Multimedia

    (2013)
  • H. Ng et al.

    Human identification based on extracted gait features

    Int. J. New Comput. Archit. Appl. (IJNCAA)

    (2011)
  • P. Arora et al.

    Gait recognition using gait Gaussian image

  • Cited by (23)

    • DF-WiSLR: Device-Free Wi-Fi-based Sign Language Recognition

      2020, Pervasive and Mobile Computing
      Citation Excerpt :

      Device-based recognition automates the recognition scheme with the deployment of sophisticated commercial devices in the sensing environment. Such methods use commercial devices like depth cameras [1], and wearable inertial or motion sensors [2–4]. However, device-based sensing methods considered to be invasive and obtrusive, hence not a preferable choice for the majority of recognition applications [5,6].

    • Human gait recognition: A systematic review

      2023, Multimedia Tools and Applications
    • Robust Motion Biomarker for Alcohol Consumption

      2022, IEEE Instrumentation and Measurement Magazine
    View all citing articles on Scopus
    View full text