Free Weight Exercises Recognition Based on Dynamic Time Warping of Acceleration Data - Intelligent Computing for Sustainable Energy and Environment - page 184

Table 4. Confusion matrices of IDTW by user-independent protocol

BC TC BP FL BR LR OP DL SCR
BC 180 0 0 0 0 0 0 0 0
TC 0 172 0 0 0 0 0 3 0
BP 0 0 167 0 0 0 11 0 0
FL 0 0 0 178 0 0 2 0 0
BR 0 0 0 0 161 0 0 16 0
LR 0 0 0 0 0 177 0 3 0
OP 0 0 14 0 0 0 166 0 0
DL 0 0 0 0 0 0 0 176 4
SCR 0 0 0 0 2 0 0 2 176

4.4 Discussion

Some exercises would impose similar acceleration responses on the hardware setting,
which is the main reason for having recognition errors. The main characteristics that
one exercise different to others is the dynamics of gravity effect and the acceleration
changing mode. Therefore, it is important for the user to keep the specified starting
status of each exercise as the motion process to improve recognition accuracy.
The BP and the OP almost have the similar starting status and motion process.
Though we can distinguish them in high accuracy by user-dependent protocol, the
performance becomes low obviously in user-independent situation, especially to the
SVM and the SDTW. The reason is that the difference between the BP and the OP to
specific user is almost uniform that it can be distinguish easily. So adding the data of
specific user to the training dataset can improve recognition performance.
The separation of acceleration data stream is one of the main factors to recognition
accuracy. When using very heavy dumbbell, the motion is very slow and shaking,
which results in very close peaks with small value. In this situation, accuracy of
separation of each repetition decreases, which may reduce recognition performance.
Nine types of dumbbell exercises are given in this paper. If more types need to be
recognized, we only need to get the reference template of new type of exercise and
add to the reference library, and don't need to train examples of old types. It is an
advantage of DTW, compared to other classifier such as SVM, HMM and ANN.

5 Conclusion and Future Work

A novel recognition method of dumbbell exercises is proposed in this paper, which is
based on the improved DTW. Compared with the conventional weight exercise
recognition method, it uses only one tri-axis accelerometer instead of two, and has
achieved better performance. A separation of acceleration data stream is proposed to
obtain signal sequences for each repetition of the exercise. The separated data stream
is deployed as the input of the DTW classifier. Experimental results show that the
recognition accuracy of the proposed DTW algorithm is 98.4%.