Research Article
Nonnegative Tensor-Based Linear Dynamical Systems for Recognizing Human Action from 3D Skeletons
Table 1
Recognition rate on the MSR-Action3D dataset based on AS1, AS2, and AS3.
| Approach | AS1 | AS2 | AS3 | Overall |
| SE3[12] | 94.34 | 83.87 | 97.43 | 91.88 |
| EigenJoints[13] | 74.5 | 76.1 | 96.4 | 83.30 |
| Ensemble TS-LSTM[14] | 95.24 | 96.43 | 100 | 97.22 |
| tLDS[6] | 96.81 | 89.14 | 98.83 | 94.85 |
| Bi-LSTM[15] | 92.72 | 84.93 | 97.89 | 91.84 |
| 2NJP-nLDS | 94.13 | 92.02 | 98.56 | 94.90 |
| 2NRBD-nLDS | 96.04 | 92.86 | 96.22 | 95.04 |
| 3NJAP-nLDS | 95.67 | 93.15 | 98.64 | 95.82 |
| 3NJAD-nLDS | 96.88 | 95.72 | 98.15 | 96.92 |
| Our approach | 97.53 | 96.5 | 98.87 | 97.63 |
|
|