Ref | Year | Model | Domain | Main objective | Remarks |
---|---|---|---|---|---|
Wang et al. (2020) | 2020 | KNN, LSM, SVM, BDM, DTW and ANN | Sensor data | Recognition of human motion detection | Data preprocessing and feature extraction required |
Doniec et al. (2002) | 2020 | CNN | Video-based data | One frame for the spatial network and ten frames of optical flow stacking are used; data augmentation is used to expand the data size | Pre-processing for features is necessary. It introduces a new representation of features and nets but does not effectively represent the spatiotemporal features in HAR. Address the overfitting issue, it requires augmentation |
Nafea et al. (2021) | 2021 | NN and SVM | Smartphone-based sensor data | Recognition of fall motion detection | Nearest neighbor provides better performance than SVM. Improved classification accuracy |
Chen et al. (2021) | 2021 | DeepConv.LSTM | Wearable sensor data | Recognition of human motion detection | Provide temporal features, improved F-score |
Mokhtari et al. (2022) | 2022 | Multi-layer CNN | Sensor-based data | Extracting spatial features using multilayer CNN | Data augmentation is required to avoid overfitting. Data preprocessing is required |
Raziani and Azimbagirad (2022) | 2022 | CNN based on grid search optimization | Wearable sensor data | Improved classification accuracy for different HAR dataset | Required data preprocessing and self-crafted feature extraction |
Dua et al. (2021) | 2021 | GAM-based deep learning methods | Wearable sensor data | Better accuracy and reduced computational cost | Restricted Boltzmann machine model, RRN, CNN is discussed for HAR task |
Gupta et al. (2021) | 2021 | MOGP-HMM-based models | Smartphone-based sensor data | Captured complex varieties using the probabilistic interval-based model and CRP model | Use in operational research and machine learning techniques |
Wang et al. (2021) | 2021 | Modified DRN model | Wearable sensor data | Modified DRN is proposed with smooth pooling layer to improve the classification accuracy | Various publicly available datasets are used to check the accuracy of the model |
Tang et al. (2020) | 2020 | Light weight deep CNN is proposed based on Lego filters | Sensor-based data | NB, SVM, and DT are used to detect various HAR activities | Data preprocessing and feature extraction required. Deep learning model used for spatial feature extraction and conventional ML models used for classification |
Tang et al. (2022) | 2022 | Create a new CNN design that incorporates statistics features | Sensor-based data | To maintain both local and global properties | CNN is used together with statistical features to improve classification accuracy |
Lu et al. (2019) | 2019 | Trajectory-pooled deep-convolutional descriptor | Video-based sensor data | Two-stream ConvNets; using improved dense trajectories (iDTs) features based on CNN for spatial feature extraction | Data preprocessing and feature extraction required |
Abdel-Salam et al. (2020) | 2021 | Hybrid CNN with LSTM | Sensor-based data | Extract spatial-temporal features from CNN and LSTM respectively | Robust nature of hybrid mechanism |
Senthilkumar et al. (2022) | 2022 | LSTM-CNN | Smartphone-based sensor video data | Extract spatial-temporal features from CNN and LSTM respectively | It was determined that the LSTM-CNN model was the best method for examining long-term activity recognition |
Wu et al. (2019) | 2019 | CNN-LSTM | Smartphone-based sensor data | Numerous models including 1D CNN, a multichannel CNN, and CNN-LSTM were improved | It was determined that the CNN-LSTM model was the best method for examining long-term activity recognition |
Yang et al. (2019) | 2019 | BILSTM-CNN Kmeans | Text data | This algorithm treats the feature extraction and clustering as a united process | The goal is to cluster texts into different clusters based on extracted temporal-spatial semantic representation |
In this paper | 2023 | Weighted BILSTM-CNN-Ensemble | Smartphone-based sensor data | The proposed scheme extracts temporal-spatial features, and the classification is done using an ensemble | The goal is to classify HAR using the effective selection of the SBS data, and the effective extraction of spatial and temporal features |