Human Activity Recognition with Noise-Injected Time-Distributed AlexNet

. 2025 Sep 11;10(9):613. doi: 10.3390/biomimetics10090613

Algorithm 1 Frame-wise feature extraction and sequence aggregation using time-distributed AlexNet.

Require: Video sequence

V = {F_{1}, F_{2}, \dots, F_{T}}

, where

F_{t}

is the

t^{th}

frame
Require: CNN layers

L_{1}, L_{2}, \dots, L_{n}

with shared weights across time

1:
for each frame $F_{t}$ in V do
2:
$x_{t} \leftarrow F_{t}$
3:
for each CNN layer $L_{i}$ in ${L_{1}, \dots, L_{n}}$ do
4:
$x_{t} \leftarrow L_{i} (x_{t})$ {Apply CNN layer with shared weights}
5:
end for
6:
$h_{t} \leftarrow Flatten (x_{t})$ {Flatten frame-wise feature map}
7:
Store $h_{t}$ into H
8:
end for
9:
$H_{concat} \leftarrow Flatten (H)$ {Flatten all time steps into one vector of shape $(T \cdot D)$ }
10:
$y \leftarrow Dense Layers (H_{concat})$ {Pass to fully connected classifier}
11:
return y {Predicted activity class}