Respiratory Motion Estimation with Hybrid Implementation - Prediction and Classification of Respiratory Motion - page 84

Table 4.3 CPU time used in the target estimation

Datasets Recording time (min) CPU time used (millisecond/# total frame)

HEKF DEKF

DB00 50.80 9.4306 7.1737

DB01 93.36 9.7759 7.2836

DB02 90.39 10.8872 7.1532

DB03 113.00 10.8578 6.9824

DB04 60.16 10.0511 7.1556

DB05 64.85 10.3541 7.3941

DB06 102.83 10.5332 7.1505

DB07 70.93 9.4372 6.7484

DB08 145.21 11.2489 7.1755

DB09 92.67 10.3379 7.0038

DB10 119.55 11.3506 7.2783

DB11 69.72 9.5831 7.0640

DB12 85.34 9.6143 6.8265

DB13 93.88 11.2510 7.4613

DB14 86.52 9.5256 7.5890

Table 4.3 shows the performance of CPU time used. As you can see in
Table 4.3 , HEKF method needs more time comparing to DEKF. We think that the
actual difference for CPU time used in Table 4.3 mainly comes from the calcu-
lation of the coupling matrix and the separate neural network for channel number.
Although 30.07 % more time is required to implement the proposed HEKF, it is a
modest tradeoff to consider the better performance than better computational time
under the condition that PC speed is improving these days.

4.5 Summary

In this Chapter we have presented respiratory motion estimation with hybrid
implementation of EKF, called HEKF. Our new method has two main contribu-
tions to improve the traditional EKF-based recurrent neural network target
tracking. The first contribution is to present a new approach to split the whole
RMLP with the complicated neuron number into a couple of RMLPs with the
simple neuron number to adjust separate input channels. The second contribution
is to comprehensively organize the multiple channel sensory process by adapting
the coupling technique using multiple channel inputs for the mutually exclusive
groups to compensate the computational accuracy.
The experiment results validated that the prediction overshoot of the proposed
HEKF was improved for 13 datasets among 15 datasets by 62.95 %. The proposed
HEKF showed the better performance by 52.40 % NRMSE improvement in the
average of the prediction time horizon. We have evaluated that a proposed HEKF