Driving Behavior Recognition Method Based on Tutor-Student Network

Jinghui Chu; Shan Zhang; Wenhao Tang; Wei Lü

doi:10.3788/LOP57.061019

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 6 >Page 061019 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 6, 061019 (2020)

Driving Behavior Recognition Method Based on Tutor-Student Network

Jinghui Chu, Shan Zhang, Wenhao Tang, and Wei Lü^*

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/LOP57.061019 Cite this Article Set citation alerts

Jinghui Chu, Shan Zhang, Wenhao Tang, Wei Lü. Driving Behavior Recognition Method Based on Tutor-Student Network[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061019 Copy Citation Text

show less

Fig. 1. Traditional teacher-student network and proposed method. (a) Traditional teacher-student network; (b) proposed method

Download full size

Fig. 2. Structural diagram of proposed model

Download full size

Fig. 3. Feature activation maps obtained from feature maps of different convolution layers in tutor network

Download full size

Fig. 4. Flow chart of guiding tailoring module

Download full size

Fig. 5. Ten kinds of action image samples from Kaggle dataset

Download full size

Fig. 6. Ten kinds of action image samples from AUC data set

Download full size

Fig. 7. Visualization results on Kaggle data set

Download full size

Model	Randomcropping	Imageresolution/PPI	Accuracy /%
ResNet18	No	224	89.30
ResNet18	Yes	224	92.56
ResNet50	No	224	90.17
ResNet50	Yes	224	94.93

Table 1. Comparative experimental results with and without data enhancement in Kaggle dataset

Model	Parameterquantity	Flops	Imageresolution/PPI	Accuracy /%
ResNet18	11,181,642	1.82G	224	92.56
ResNet18	11,181,642	7.28G	448	94.21
ResNet50	23,528,522	4.12G	224	94.93
ResNet50	23,528,522	16.47G	448	96.48

Table 2. Experimental results of setting different resolutions in Kaggle data set

Model	Parameterquantity	Flops	Imageresolution/PPI	Accuracy /%
ResNet18	11,181,642	1.82G	224	92.56
ResNet34	21,289,802	3.67G	224	94.67
ResNet50	23,528,522	4.12G	224	94.93
ResNet101	42,520,650	7.84G	224	95.69
S-Net(ResNet18)	22,363,284	9.10G	224	92.78
S-Net(ResNet50)	34,710,164	11.40G	224	96.29

Table 3. Comparative experimental results of different models in Kaggle data set

Model	Parameter quantity	Flops	Image resolution/PPI	Accuracy /%
S-Net (ResNet18)	22,363,284	9.10G	224	92.78
S-Net (ResNet50)	34,710,164	11.40G	224	96.29
ResNet18+ResNet50(ensemble)	34,710,164	5.94G	224(T-Net)+224(S-Net)	95.95
ResNet18+ ResNet50(ensemble)	34,710,164	11.40G	448(T-Net)+224(S-Net)	96.10
T-Net(ResNet18)+S-Net(ResNet18)	22,363,284	9.10G	448(T-Net)+224(S-Net)	95.99
T-Net(ResNet18)+S-Net(ResNet50)	34,710,164	5.94G	224(T-Net)+224(S-Net)	96.56
T-Net(ResNet18)+S-Net(ResNet50)	34,710,164	11.40G	448(T-Net)+224(S-Net)	97.92

Table 4. Comparative experimental results of model joint discrimination in Kaggle data set

Model	Source	Accuracy /%
AlexNet^[16]	Original	93.65
AlexNet^[16]	Skin segmented	93.60
AlexNet^[16]	Face	84.28
AlexNet^[16]	Hands	89.52
AlexNet^[16]	Face+hands	86.68
Inception V3^[16]	Original	95.17
Inception V3^[16]	Skin segmented	94.57
Inception V3^[16]	Face	88.82
Inception V3^[16]	Hands	91.62
Inception V3^[16]	Face+hands	90.88
ResNet50	Original	94.87
S-Net(ResNet50)	Original	95.20
T-Net(ResNet18)+S-Net(ResNet50)	Original	95.71

Table 5. Model comparison results on AUC dataset

Jinghui Chu, Shan Zhang, Wenhao Tang, Wei Lü. Driving Behavior Recognition Method Based on Tutor-Student Network[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061019

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information