Author Affiliations
School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou, Gansu 730070, Chinashow less
Fig. 1. CNN based on single-column deep spatiotemporal counting
Fig. 2. Structure of the improved FCN
Fig. 3. Structure of the ST counting network
Fig. 4. Counting results of our model on the UCSD data set
Fig. 5. Experimental results of our model on the Mall data set. (a) Density map; (b) counting result
Fig. 6. Experimental results of our model on the self-built data set. (a) Density map; (b) counting result
Fig. 7. Accuracies of different models
Fig. 8. Training loss curves of the network before and after the improvement
Models | fMAE | fMSE |
---|
ConvLSTM[3] | 1.30 | 1.79 | Bidirectional ConvLSTM[3] | 1.13 | 1.43 | Ref. [19] | 2.24 | 7.97 | Ref. [20] | 2.25 | 7.82 | Ref. [10] | 1.54 | 3.02 | Ref. [21] | 2.07 | 6.86 | Ours | 1.05 | 1.59 |
|
Table 1. Performance indexes of different models on the UCSD data set
Model | fMAE | fMSE |
---|
ConvLSTM[3] | 2.24 | 8.50 | Bidirectional ConvLSTM[3] | 2.10 | 7.60 | Ref. [20] | 3.59 | 19.00 | Ref.[21] | 3.43 | 17.70 | Ours | 1.95 | 7.50 |
|
Table 2. Performance indexes of different models on the Mall data set
Model | fMAE | fMSE |
---|
ConvLSTM[3] | 4.51 | 5.91 | MCNN[22] | 3.81 | 4.92 | Ours without ST | 4.32 | 5.21 | Ours | 3.51 | 5.10 |
|
Table 3. Performance indexes of different models on self-built data set
Data set | No dilated No ST | dilated+No ST |
---|
fMAE | fMSE | fMAE | fMSE |
---|
UCSD | 1.71 | 4.25 | 1.52 | 4.13 | Mall | 2.89 | 9.01 | 2.13 | 8.51 | Self-built | 4.74 | 6.65 | 4.32 | 5.21 |
|
Table 4. Confirmation experiment results1 of different data sets
Data set | No dilated No ST | No dilated+ ST | dilated+ST |
---|
fMAE | fMSE | fMAE | fMSE | fMAE | fMSE |
---|
UCSD | 1.71 | 4.25 | 1.41 | 3.52 | 1.05 | 1.59 | Mall | 2.89 | 9.01 | 2.01 | 8.43 | 1.95 | 7.50 | Self-built | 4.74 | 6.65 | 4.33 | 5.23 | 3.51 | 5.10 |
|
Table 5. Confirmation experiment results 2 of different data sets