Vision Transformer and Deep Sequence Learning for Human Activity Recognition in Surveillance Videos

Table 3

The proposed LSTM network to capture long-range temporal information from video sequences.

Layer (type)	Output shape	No. of parameters
Input data	(None, 30, 1000)	0
LSTM	(None, 30, 128)	578048
LSTM	(None, 64)	49408
Dropout	(None, 64)	0
Batch normalization	(None, 64)	256
Activation	(None, 64)	0
Dense	(None, 64)	4160
Dense	(None, 51)	3315
Activation	(None, 51)	0