Fig. 1. Flow chart of building radar time-frequency image data set
Fig. 2. Radar time-frequency spectrograms of radar for six human activities. (a) Walking; (b) boxing; (c) crawling on the ground; (d) diving; (e) standing forward and jumping; (f) running
Fig. 3. Structural diagram of SE module
Fig. 4. Multi-task recognition model structure based on CNN
Fig. 5. Diagram of joint loss function generation
Fig. 6. Loss and accuracy curves. (a) Loss curves; (b) accuracy curves
Crossvalidationset No. | Recognitionrate of MR /% | Recognitionrate of SR /% |
---|
HA | HD | HT | HA | HD | HT |
---|
1 | 99.94 | 98.78 | 98.74 | 99.70 | 95.99 | 95.72 | 2 | 99.94 | 98.80 | 98.77 | 99.88 | 96.70 | 96.93 | 3 | 99.97 | 98.90 | 98.89 | 99.90 | 97.91 | 97.84 | 4 | 100.00 | 99.93 | 99.93 | 99.96 | 99.37 | 99.34 | Mean | 99.96 | 99.10 | 99.08 | 99.86 | 97.49 | 97.38 |
|
Table 1. Results of multi-task and single-task recognitions
Model | Recognition rate /% |
---|
HA | HD | HT |
---|
D-MSSE | 99.98 | 99.62 | 99.61 | D-MS | 100 | 99.88 | 99.88 | D-SE | 99.99 | 90.47 | 90.47 |
|
Table 2. Recognition results of different network structures
Loss | Recognition rate /% |
---|
HA | HD | HT |
---|
LE | 99.99 | 99.74 | 99.73 | LE +LM | 99.99 | 99.86 | 99.86 | LE+LC | 100 | 99.87 | 99.86 |
|
Table 3. Recognition results of different loss optimization
L /s | Recognition rate /% |
---|
HA | HD | HT |
---|
0.2 | 97.99 | 78.59 | 77.68 | 0.4 | 99.75 | 89.25 | 89.13 | 0.6 | 99.91 | 92.47 | 92.44 | 0.8 | 99.98 | 99.73 | 99.71 |
|
Table 4. Recognition results under different STFT window lengths
NSR /dB | Recognition rate /% |
---|
HA | HD | HT |
---|
-20 | 67.53 | 24.4 | 19.06 | -10 | 96.91 | 77.63 | 76.49 | 0 | 99.91 | 90.8 | 90.75 | 10 | 99.99 | 99.5 | 99.49 | +∞ | 100 | 99.93 | 99.93 |
|
Table 5. Recognition results under different NSR
Data size | Recognition rate /% |
---|
HA | HD | HT |
---|
400 | 99.94 | 97.52 | 97.46 | 600 | 99.97 | 99.71 | 99.57 | 800 | 99.97 | 99.62 | 99.59 | 1000 | 99.99 | 99.76 | 99.75 | 1300 | 99.99 | 99.85 | 99.84 | 1500 | 100 | 99.93 | 99.93 |
|
Table 6. Recognition results under different training data sizes
Algorithm | Recognition rate /% |
---|
HA | HD | HT |
---|
PRGC | 99.94 | 98.9 | 98.84 | MRN | 99.97 | 99.9 | 99.85. | MCNN | 100 | 99.93 | 99.93 |
|
Table 7. Comparison of results of algorithms