Deep driver behavior detection model based on human brain consolidated learning for shared autonomy systems

•Images of different size are used to train to improve the accuracy of DBD.•Multiple DBD models are trained respectively based on the transfer learning.•A model named Consolidation Training (CT) is trained based on weight data with DBD.•Mish with no boundary and good smoothness is used as the activa...

Full description

Saved in:
Bibliographic Details
Published in:Measurement : journal of the International Measurement Confederation 2021-07, Vol.179, p.109463, Article 109463
Main Authors: Huang, Tao, Fu, Rui, Chen, Yunxing
Format: Article
Language:English
Subjects:
Anthropomorphism
Automobile drivers
Automobile driving
Autonomy
Brain
Consolidation
Deep neural network
Driver behavior
Learning
Model accuracy
Sensors
Shared autonomy
Traffic accidents & safety
Transfer learning, consolidated learning
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•Images of different size are used to train to improve the accuracy of DBD.•Multiple DBD models are trained respectively based on the transfer learning.•A model named Consolidation Training (CT) is trained based on weight data with DBD.•Mish with no boundary and good smoothness is used as the activation function.•A novel visualization method of the attention area is proposed. To achieve efficient shared autonomy, driver behavior detection (DBD) is undoubtedly required. This paper investigates a deep driver behavior detection (DDBD) model. To overcome the low accuracy of DBD due to a lack of driver behavior data, the similarity of some driver behavior characteristics, and the ignorance of multi-scale structure and texture information, a DDBD model based on human brain consolidated learning (HBCL) is proposed. First, multiple DBD models with input information of different scales are trained based on transfer learning. Then, a new model called consolidation training (CT) using the Mish is trained based on the weight data from the first step. Finally, a novel method for the visualization of the attention area is proposed. The experimental results demonstrate that the proposed model achieved the highest accuracy (94.72% on the Kaggle-driving test dataset), generalization and real-time performance, the attention area is more anthropomorphic as compared with existing state-of-the-art models.
ISSN:0263-2241
1873-412X
DOI:10.1016/j.measurement.2021.109463