A Stacked Human Activity Recognition Model Based on Parallel Recurrent Network and Time Series Evidence Theory

. 2020 Jul 19;20(14):4016. doi: 10.3390/s20144016

Algorithm 1 Time series evidence combination rule

Require:
Revised evidence ${\tilde{E}}_{s_{i}}^{t}$ ; the hyper-parameters $N_{i}$ ; the number of classifiers/activities n; current time interval T;
Ensure:
The final combination result $E_{f}$ ;
1:
Initial the weights $W^{P}$ and ${\hat{W}}^{P}$ ;
2:
Initial $E_{f} = [0, 0, \dots, 0, 1]$ ;
3:
for $i = 1$ ; $i \leq n$ ; $i + +$ do
4:
for $t = T$ ; $t \geq T - N_{i}$ ; $t - -$ do
5:
Calculate each ${\hat{E}}_{s_{i}}^{t}$ of classifier $s_{i}$ in time interval t with $W^{P}$ by Equations (7) and (8);
6:
end for
7:
Calculate each intra-classifier combination result ${\hat{E}}_{s_{i}}$ of classifier $s_{i}$ by Equation (6);
8:
$E_{f} \leftarrow E_{f} \oplus {\hat{W}}^{P} ⊙ {\hat{E}}_{s_{i}}$ by Equations (6)–(8);
9:
end for
10:
if train model then: Minimize the loss function and calculate the change of weights $Δ W^{P}$ and $Δ {\hat{W}}^{P}$ ; $W^{P} \leftarrow W^{P} + Δ W^{P}$ ; ${\hat{W}}^{P} \leftarrow {\hat{W}}^{P} + Δ {\hat{W}}^{P}$ ;
11:
else if test model then: Output $W^{P}$ and ${\hat{W}}^{P}$ ;
12:
end if