Table 4.
Intrasubject gesture recognition accuracy in comparison with related works on five databases.
| Machine learning (ML) model | Type of ML model | Input of ML model | Database | Num. of gestures for classification | Window length | ||||
| 50 ms | 100 ms | 150 ms | 200 ms | Trial | |||||
|
| |||||||||
| Random forests [13] | Shallow learning | 5 hand-crafted features | NinaProDB1 | 50 | N.A. | N.A. | N.A. | 75.3% | N.A. |
| Dictionary learning [36] | Shallow learning | MLSVD-based features | NinaProDB1 | 52 | N.A. | N.A. | N.A. | N.A. | 97.4% |
| HuNet [26] | CNN-RNN | Phinyomark feature set | NinaProDB1 | 52 | N.A. | N.A. | 86.8% | 87.0% | 97.3% |
| MV-CNN [31] | Multiview CNN | 3 views of sEMG | NinaProDB1 | 52 | 85.8% | 86.8% | 87.4% | 88.2% | N.A. |
| ChengNet [49] | CNN | Multi-sEMG-features image | NinaProDB1 | 52 | N.A. | N.A. | N.A. | 82.5% | N.A. |
| HVPN-2-view | Multi-view CNN | 2 views of sEMG | NinaProDB1 | 52 | 85.4% | 86.5% | 87.2% | 88.1% | 97.8% |
| HVPN | Multi-view CNN | Same as [31] | NinaProDB1 | 52 | 86.0% | 86.9% | 87.7% | 88.4% | 98.0% |
|
| |||||||||
| Random forests [13] | Shallow learning | Hand-crafted features | NinaProDB2 | 50 | N.A. | N.A. | N.A. | 75.3% | N.A. |
| ZhaiNet [25] | CNN | sEMG spectrogram | NinaProDB2 | 50 | N.A. | N.A. | N.A. | 78.7% | N.A. |
| HuNet [26] | CNN-RNN | Phinyomark feature set | NinaProDB2 | 50 | N.A. | N.A. | N.A. | 82.2% | 97.6% |
| MV-CNN [31] | Multiview CNN | 3 views of sEMG | NinaProDB2 | 50 | 80.6% | 81.1% | 82.7% | 83.7% | N.A. |
| HVPN-2-view | Multiview CNN | 2 views of sEMG | NinaProDB2 | 50 | 82.7% | 83.8% | 83.3% | 85.0% | 97.8% |
| HVPN | Multiview CNN | Same as [31] | NinaProDB2 | 50 | 82.3% | 84.1% | 84.8% | 85.8% | 98.1% |
|
| |||||||||
| Support vector machine (SVM) [13] | Shallow learning | 5 hand-crafted features | NinaProDB3 | 50 | N.A. | N.A. | N.A. | 46.3% | N.A. |
| MV-CNN [31] | Multiview CNN | 3 views of sEMG | NinaProDB3 | 50 | N.A. | N.A. | N.A. | 64.3% | N.A. |
| ED-TCN [27] | TCN | MAV sequences | NinaProDB3 | 41 | N.A. | N.A. | 63.5% | N.A. | N.A. |
| HVPN-2-view | Multiview CNN | 2 views of sEMG | NinaProDB3 | 50 | 64.4% | 65.7% | 66.8% | 67.9% | 80.3% |
| HVPN | Multiview CNN | Same as [31] | NinaProDB3 | 50 | 64.5% | 65.9% | 66.9% | 68.2% | 80.7% |
|
| |||||||||
| Random forests [37] | Shallow learning | mDWT features | NinaProDB4 | 53 | N.A. | N.A. | N.A. | 69.1% | N.A. |
| MV-CNN [31] | Multiview CNN | 3 views of sEMG | NinaProDB4 | 53 | N.A. | N.A. | N.A. | 54.3% | N.A. |
| HVPN-2-view | Multiview CNN | 2 views of sEMG | NinaProDB4 | 53 | 60.1% | 63.2% | 67.6% | 72.1% | 81.1% |
| HVPN | Multiview CNN | Same as [31] | NinaProDB4 | 53 | 58.3% | 67.1% | 70.5% | 72.9% | 81.7% |
|
| |||||||||
| SVM [37] | Shallow learning | mDWT features | NinaProDB5 | 41 | N.A. | N.A. | N.A. | 69.0% | N.A. |
| ShenNet [50] | Stacking-based CNN | TD, FD and TFD features | NinaProDB5 | 40 | N.A. | N.A. | N.A. | 72.1% | N.A. |
| MV-CNN [31] | Multiview CNN | 3 views of sEMG | NinaProDB5 | 41 | N.A. | N.A. | N.A. | 90.0% | N.A. |
| HVPN-2-view | Multiview CNN | 2 views of sEMG | NinaProDB5 | 41 | 88.7% | 89.1% | 89.9% | 90.1% | 98.8% |
| HVPN | Multiview CNN | Same as [31] | NinaProDB5 | 41 | 88.7% | 89.3% | 90.0% | 90.3% | 98.4% |
N.A. denotes not applicable, and bold entries indicate our proposed method. HVPN-2-view refers to the proposed HVPN framework with the “two-view” configuration (i.e., using v1 and v2 as its input). †It should be mentioned that existing MCIs seldom segment raw sEMG signals by trial due to the constraint that the maximal response time of an MCI should be kept below 300 ms [40, 41]. ‡For experiments on HVPN, the predicted class label of each gesture trial is obtained by majority voting on all 200 ms sliding windows within it.