1530-437X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2017.2678484, IEEE Sensors Journal Saho et al.: GAIT CLASSIFICATION OF YOUNG ADULTS, ELDERLY NON-FALLERS AND FALLERS 1 Abstract— This letter presents a gait classification technique for identification of individuals with different gait patterns using simulated micro-Doppler radar remote sensing data. Proposed feature parameters for the classification are principal components of velocities extracted via micro-Doppler radar signals generated using motion capture-based kinematic data. Distinct differences were found in the proposed parameters among three groups of subjects with different gait patterns: healthy young and elderly adults, and elderly adults with a history of falls (elderly fallers). Index Terms—micro-Doppler radar, gait classification, falls, elderly, principal component analysis I. INTRODUCTION simple remote-sensing technique to identify individuals with a high risk of falls is a necessary part of health monitoring for prevention of fall-related injuries in elderly people. Since falls most frequently occur during walking, many studies have investigated age-related changes in balance control during gait to enhance fall risk assessment [1]–[3]. Although optical motion capture (Mocap) systems have been traditionally used for this type of study [1], [2], their limited portability, and the time-consuming marker placement and data analysis required for their use, preclude their general use. An accelerometry-based approach could be an alternative [3], but would still require that subjects wear accelerometers. Micro-Doppler radar (MDR) allows remote-sensing of human motion without the need for sensors to be worn by the subject [4], and its effectiveness for human motion classification has been verified [5]. Although MDR has been used to detect fall events [6], MDR can also be used to classify individuals with different gait patterns. This would allow us to detect individuals who are prone to a fall before it happens. In this study, we used a gait classification technique based on MDR signals to identify individuals with different gait patterns. Manuscript received September 6, 2016. This work was supported in part by the Ministry of Internal Affairs and Communications of Japan and JSPS KAKENHI 16K16093. K. Saho and M. Masugi are with the College of Science and Engineering, Ritsumeikan University, Shiga 525-8577, Japan (phone: +81-77-561-3483; e-mail: saho@fc.ritsumei.ac.jp, masugi@fc.ritsumei.ac.jp). M. Fujimoto is with the College of Sport and Health Science, Ritsumeikan University, Shiga 525-8577, Japan (e-mail: mfujimo@fc.ritsumei.ac.jp). L.-S. Chou is with the Department of Human Physiology, University of Oregon, Eugene, OR 97403, United States (e-mail: chou@uoregon.edu). Velocity parameters were extracted from MDR signals, which were generated using Mocap-based kinematic data from three subject groups while walking: healthy young and elderly adults, and elderly adults with a history of falls. We verified that the principal components of the extracted parameters clearly differed among three groups. II. TESTED DATA AND RADAR SIGNAL GENERATION Realistic MDR signals were generated using the Mocap data previously presented in [1]. The accuracy of MDR signals generated from Mocap data has been previously demonstrated [5], and the use of Mocap data to generate simulated MDR data allows initial evaluation of clinical applications for MDR without requiring time-consuming and costly MDR experiments. Study subjects included 15 healthy young adults [Young: 7 men; mean age 22.1 ± 1.9 years, mean height 170.4 ± 11.0 cm], 15 healthy elderly adults [Elderly: 6 men; mean age 70.0 ± 3.2 years, mean height 170.1 ± 8.7 cm], and 15 elderly adults with a history of falls [Fallers: 3 men; mean age 71.9 ± 4.3 years, mean height 164.2 ± 8.6 cm]. Subjects were instructed to walk barefoot at a self-selected comfortable pace along a 10-m walkway. Other details have been reported in [1]. Fig. 1 depicts an assumed measurement situation and MDR parameters. Received signals were generated based on the ray-tracing process [4], [5] using interpolated Mocap data. White Gaussian noises were added to the generated signals at a minimum signal-to-noise ratio of 20.0 dB. III. MDR SIGNAL PROCESSING PROCEDURE Short-time Fourier transformations of the MDR signals were performed to acquire spectrograms (time and radial velocity distribution) [4]–[6]. Fig. 2 depicts a representative spectrogram from the Young group, where the hamming window function was used with a length of 128 samples (72.5 ms). Details of the MDR signal processing are reported in [4]. Velocity parameters were then extracted from the spectrograms. Similar to [4] and [5], the upper envelope vu(t) was extracted as the significant peak corresponding to the maximum velocity at each time, and power-weighted mean velocity vm(t) was also extracted (Fig. 2). vu(t) and vm(t) correspond to leg and torso speeds, respectively [4]. Using these, the following parameters are calculated: vmean = Gait Classification of Young Adults, Elderly Non-Fallers and Elderly Fallers Using Micro- Doppler Radar Signals: A Simulation Study Kenshi Saho, Member IEEE, Masahiro Fujimoto, Masao Masugi, and Li-Shan Chou A