Lite-3DCNN Combined with Attention Mechanism for Complex Human Movement Recognition

Table 1

The network structure and parameters of this framework (20 class).

Layers	Output shape	Parameters
Input layer	32, 32, 20, 3	0
conv3d	32, 32, 20, 32	2624
activation	32, 32, 20, 32	655360
conv3d_1	32, 32, 20, 32	27680
activation_1	32, 32, 20, 32	0
max_pooling3d	10, 10, 6, 32	0
Dropout	10, 10, 6, 32	0
conv3d_2	10, 10, 6, 64	55360
activation_2	10, 10, 6, 64	0
conv3d_3	10, 10, 6, 64	110656
activation_3	10, 10, 6, 64	0
max_pooling3d_1	3, 3, 2, 64	0
dropout_1	3, 3, 2, 64	0
time_distributed (flatter)	3, 384	0
self__attention	3, 512	589824
Dense	3, 512	262656
batch_normalization	3, 512	2048
dropout_2	3, 512	0
global_average_pooling1d	512	0
dense_1	20	10260