CNN-LSTM model for exteremely low-resolution FIR sequences
Accuracy (LOSOCV): 96.98%
| walk | sitdown | standup | falling | no action | |
|---|---|---|---|---|---|
| walk | 349 | 6 | 4 | 0 | 1 |
| sitdown | 3 | 327 | 10 | 3 | 17 |
| standup | 2 | 3 | 343 | 0 | 12 |
| falling | 2 | 4 | 0 | 353 | 1 |
| no action | 4 | 4 | 0 | 0 | 1072 |
Accuracy (LOSOCV): 90.87%
| walk | sitdown | standup | falling | sitting | lying | standing | |
|---|---|---|---|---|---|---|---|
| walk | 355 | 1 | 2 | 0 | 0 | 1 | 1 |
| sitdown | 1 | 310 | 39 | 7 | 2 | 0 | 1 |
| standup | 1 | 37 | 318 | 1 | 0 | 0 | 3 |
| falling | 0 | 3 | 4 | 349 | 1 | 3 | 0 |
| sitting | 0 | 2 | 1 | 0 | 299 | 9 | 49 |
| lying | 0 | 0 | 0 | 6 | 5 | 346 | 3 |
| standing | 0 | 0 | 6 | 0 | 40 | 1 | 313 |
To do:
- log confusion matrix and accuracy for the branches
- experiments on 5 & 7 classess
- check out cosine loss (negative/neutral impact on accuracy)
- pretraining
- train the streams separately