TY - JOUR
T1 - HCF: A Hybrid CNN Framework for Behavior Detection of Distracted Drivers
AU - Huang, Chen
AU - Huang, Chen
AU - Wang, Xiaochen
AU - Cao, Jiannong
AU - Wang, Shihui
AU - Wang, Shihui
AU - Zhang, Yan
AU - Zhang, Yan
N1 - Funding Information:
This work was supported in part by the National Natural Science Foundation of China under Grant 61977021, in part by the Natural Science Foundation of Hubei Province under Grant 2018CFB692, in part by the Science and Technology Innovation Program of Hubei Province under Grant 2018ACA13, and in part by the Science and Technology Innovation Major Program of Hubei Province under Grant 2019ACA144.
Publisher Copyright:
© 2013 IEEE.
Copyright:
Copyright 2020 Elsevier B.V., All rights reserved.
PY - 2020
Y1 - 2020
N2 - Distracted driving causes a large number of traffic accident fatalities and is becoming an increasingly important issue in recent research on traffic safety. Gesture patterns are less distinguishable in vehicles due to in-vehicle physical constraints and body occlusions from the drivers. However, by capitalizing on modern camera technology, convolutional neural network (CNN) can be used for visual analysis. In this paper, we present a hybrid CNN framework (HCF) to detect the behaviors of distracted drivers by using deep learning to process image features. To improve the accuracy of the driving activity detection system, we first apply a cooperative pretrained model that combines ResNet50, Inception V3 and Xception to extract driver behavior features based on transfer learning. Second, because the features extracted by pretrained models are independent, we concatenate the extracted features to obtain comprehensive information. Finally, we train the fully connected layers of the HCF to filter out anomalies and hand movements associated with non-distracted driving. We apply an improved dropout algorithm to prevent the proposed HCF from overfitting to the training data. During the evaluation, we apply the class activation mapping (CAM) technique to highlight the feature area involving ten tested classes of typical distracted driving behaviors. The experimental results show that the proposed HCF achieves the classification accuracy of 96.74% when detecting distracted driving behaviors, demonstrating that it can potentially help drivers maintain safe driving habits.
AB - Distracted driving causes a large number of traffic accident fatalities and is becoming an increasingly important issue in recent research on traffic safety. Gesture patterns are less distinguishable in vehicles due to in-vehicle physical constraints and body occlusions from the drivers. However, by capitalizing on modern camera technology, convolutional neural network (CNN) can be used for visual analysis. In this paper, we present a hybrid CNN framework (HCF) to detect the behaviors of distracted drivers by using deep learning to process image features. To improve the accuracy of the driving activity detection system, we first apply a cooperative pretrained model that combines ResNet50, Inception V3 and Xception to extract driver behavior features based on transfer learning. Second, because the features extracted by pretrained models are independent, we concatenate the extracted features to obtain comprehensive information. Finally, we train the fully connected layers of the HCF to filter out anomalies and hand movements associated with non-distracted driving. We apply an improved dropout algorithm to prevent the proposed HCF from overfitting to the training data. During the evaluation, we apply the class activation mapping (CAM) technique to highlight the feature area involving ten tested classes of typical distracted driving behaviors. The experimental results show that the proposed HCF achieves the classification accuracy of 96.74% when detecting distracted driving behaviors, demonstrating that it can potentially help drivers maintain safe driving habits.
KW - convolutional neural network
KW - Distracted drivers
KW - fusion model
KW - transfer learning
UR - http://www.scopus.com/inward/record.url?scp=85087333948&partnerID=8YFLogxK
U2 - 10.1109/ACCESS.2020.3001159
DO - 10.1109/ACCESS.2020.3001159
M3 - Journal article
AN - SCOPUS:85087333948
SN - 2169-3536
VL - 8
SP - 109335
EP - 109349
JO - IEEE Access
JF - IEEE Access
M1 - 9113267
ER -