Driving Behavior Primitive Classification Using CNN-Based Fusion Models

• Published in: IEEE access Vol. 12; pp. 56344 - 56355
• Main Authors: Cui, Xiaotong, Li, Xiansheng, Zheng, Xuelian, Ren, Yuanyuan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Advanced driver assistance systems; Behavioral sciences; Brain modeling; Classification; CNN-based fusion model; Convolutional neural networks; Couplings; Data mining; Deep learning; Driving behavior analysis; driving behavior primitive classification; Feature extraction; Hidden Markov models; information fusion; Model accuracy; primitive feature matrix; Semantics; Statistical methods; Vehicle driving
• DOI: 10.1109/ACCESS.2024.3391170
• ISSN: 2169-3536

Driving behavior primitives play a crucial role in semantic explanation of driving behaviors. Although much work has been done on exacting driving behavior primitives from naturalistic driving data, few studies was published on primitive classification. Driving behavior primitives are typically described by multi-dimensional variables with varying durations, which leads to the inefficiency of the traditional classification methods. There hence, a CNN-based fusion model for primitive classification is proposed in this paper. Primitive feature matrix is constructed using statistical methods for the four basic and the four constructed variables, which serves as the input. A 1D-CNN is employed to extract global information of the total eight variables in the feature matrix, while a 2D-CNN is used to extract the local information. The 1D-CNN and the 2D-CNN are fused in parallel using a new fusion method to combine different types of information, and two models, namely the FC-before fusion model and the FC-after fusion model, are acquired. Compared with the classical methods, the empirical results demonstrate that CNN-based fusion model can recognize driving behavior primitives more accurately. Specifically, the FC-after fusion model achieves an accuracy of 91.12% and a macro F1-score of 90.88%, while the accuracy and macro F1-score of the FC-before fusion model are 93.47% and 92.57%, respectively.