Information Technology Reference
In-Depth Information
recognition accuracy of 9 types of exercises is around 90% using Hidden Markov
Models classifiers. A RFID based system for monitoring free weight exercises was
proposed in [5]. Gaussian model-based classifiers and two tri-axis accelerometers
were deployed in [6] to recognize 16 gym exercises, which included 11 weight lifting
exercises and the average precision is 92%. However, two accelerometers are not only
high-cost, but also inconvenient to wear.
Dynamic time warping (DTW) is one of the commonly used methods for similarity
measurement in the time series classification [7], which is widely used in speech
recognition, medical analysis, and moving object identification, e.g. multi-motion
recognition based on accelerometer [8][13]. Recently, much work has been done to
improve DTW accuracy [9-11] and to reduce computation time [13].
In this paper, a single tri-axis accelerometer is attached to the lifting hand glove to
recognize nine weight lifting exercises. Improved DTW is adopted to classify
exercises types by calculating the similarity between reference template series and test
series. Acceleration series data in each repetition is separated by peak analysis and
used as the input to 3D DTW classifiers. Experimental results revealed that the
proposed method can be applied to various weight exercises. Exercise recognition
accuracy is 98% for the user dependent case, and the accuracy is around 95% for
user-independent cases.
The remainder of this paper is organized as follows. Section 2 describes a
recognition method for weight lifting exercises. Section 3 presents the development of
the recognition algorithms. Experimental results and analysis are given in Section 4 to
show the feasibility and performance of the proposed algorithm. Finally, a brief
conclusion and future work are presented in Section 5.
2
Method of Weight Exercises Recognition
2.1
Response of Every Type of Exercises
Nine common weight lifting exercises are adopted: Biceps curl(BP), Triceps
curl(TC), Bench press(BP), Fly(FL), Bent-over row(BR), Literal raise(LR), Overhead
dumbbell press(OP), Deadlift(DL), Stand calf raises(SCR). Posture of each type can
be found in [4].
A MTx 3-DOF orientation tracker (produced by Xsens) is used in the experiment,
as in Fig. 1. Each MTx unit has a tri-axial accelerometer, a tri-axial gyroscope, and a
tri-axial magnetometer. But only the accelerometer is used in this experiment. The
tracker is installed on back of the right hand glove. As in Fig. 1, with the palm down,
the fingers pointing the x-axis, and y-axis at the left, and the z-axis is up.
Measurement range of acceleration is ±2g. Fig. 2 shows the curve of acceleration data
filtered by 5th order Butterworth low pass filter with sampling rate of 100Hz. Most
types of exercises are quite distinct from each other, but BP and OP are very similar
as they have the same motion path. The only difference is that body posture of BP is
lying, and OP is standing.
Search WWH ::
Custom Search