Strike Activity Detection and Recognition Using Inertial Measurement Unit towards Kendo Skill Improvement Support System

Abstract

In the field of sports, there are increasing opportunities to use inertial measurement units (IMUs) to enhance the training process and improve the performance of athletes. We focus on kendo, a traditional martial art using shinai (bamboo swords) in Japan, and propose methods for detecting and recognizing strike activities using IMUs towards realizing a kendo skill improvement support system. We used a sensor data set of strike activities obtained from 14 participants (seven kendo-experienced and seven inexperienced persons). We attached four IMUs to the participants' right wrist, waist, and shinai (tsuba and saki-gawa). First, to detect the strike activity, we calculated the dynamic time warping (DTW) distance between the training data and the time series data, and detected the strike activity sections. The proposed method detected strike activities with a high accuracy of 95.0%. Next, to recognize the strike activity, we recognized five types (Center-Men, Right-Men, Left-Men, Dō, and Kote). In the person-dependent (PD) case, we achieved an accuracy of 89.5% using data of the right wrist. In the person-independent (PI) case, we achieved an accuracy of 54.9% using IMUs attached to the three positions. These results clarified the points to be improved in the proposed method to realize the support system.

Inertial Measurement Unit-sensor-based Short Stick Exercise Tracking to Improve Health of Elderly People

“…Here, the sampling rate is 100 Hz. In this study, data preprocessing and feature extraction are performed in accordance with the findings of existing studies; (10,15) we apply a sliding window with a window size of 1.28 s (1 window = 128 samples) and a step size of 0.64 s. We first apply a noise reduction process to nine different time-series data using a 0.3 s median filter. Next, we separate the sensor data into time-domain and frequency-domain series using a Fast Fourier transform.…”

“…They analyzed inertial sensor data of striking motions of experienced and inexperienced kendo players and reported that the proposed method based on dynamic-time warping could detect striking motions with an F value of 89.9%. (10) Blank et al attached inertial sensors to table tennis rackets and collected data on eight basic stroke types from 10 amateur and 10 professional players. First, a single stroke was detected using an event detection algorithm.…”

ZEL: Net-Zero-Energy Lifelogging System using Heterogeneous Energy Harvesters

“…In order to apply time-related features, each data type is divided by a fixed-length window. In our system, we selected a window size of 1.24 seconds [4], [26], which is commonly used in accelerometer activity recognition. The window overlap rate is 50%.…”

“…2) Feature extraction: Since time-series data are the target of this study, we use the following 17 features that have been validated in previous studies [4], [26] related to activity recognition using accelerometers: mean, standard deviation, median absolute deviation, maximum, minimum, sum of squares, entropy, interquartile range, fourth-order Burg autoregressive model coefficients, range of minimum and maximum values, root mean square, frequency signal skewness, frequency signal kurtosis, maximum frequency component, frequency signal weighted average, frequency band spectral energy, and power spectral density.…”