Table 1.

Summary of HAR-related work presented in this section.

Work	Used Methods	Limitations
Ignatov [22], 2018	CNN + Statistical Features	Statistical feature extraction requires additional computational cost
Xia et al. [23], 2020	CNN + LSTM	The model depth and layer diversity increases the model complexity
Nafea et al. [24], 2021	CNN + BiLSTM
Yin et al. [25], 2022	CNN + BiLSTM + Attention	LSTM and GRU RNNs suffer from increased computation time, limiting their applicability to edge inference
Tan et al. [26], 2022	Conv1D + GRU + Ensemble learning
Pushpalatha and Math [27], 2022	CNN + GRU+ FC	Models tested on a single dataset do not establish the model generalization capabilities
Sikder et al. [28], 2019	CNN	Using such a DNN increases the computational cost of the model
Luwe et al. [29], 2022	CNN + BiLSTM	Using a DNN model with hybrid layers increases the model complexity and computational cost of the proposed classifier
Ronald et al. [30], 2021	CNN + BiLSTM + Inception + ResNet	Such a deep model is not the best fit for edge inference, which requires smaller models with a reduced computational cost.
Sannara EK [32], 2022	CNN + Transformer	The number of parameters is greater than 1 million
Tang et al. [33], 2021	Teacher-Student CNN
Rahimi Taghanaki et al. [34], 2021	CNN + FC + Transfer Learning	Results achieved by self-supervised and semisupervised models fall behind their supervised learning counterparts by a considerable margin
Taghanaki et al. [35], 2022	CNN + STFT + Transfer Learning