Driving Safety Assessment via Deep Learning Models on CAN-Bus Sensor Streams

Nurbakyt Agybetov, Beibut Amirgaliyev

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1016/j.procs.2026.08.417
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

This paper proposes a deep learning framework for continuous driving safety assessment based on CAN-derived vehicle dynamics data. Despite recent advances in driver behavior analysis, existing approaches often rely on window-level evaluation protocols that may introduce information leakage and primarily focus on discrete Classification, without providing continuous safety measures that reflect temporal variability in driving behavior. To address these limitations, a window-based approach is introduced, in which continuous sensor signals are segmented into overlapping temporal windows and analyzed using neural network models, including 1D convolutional neural networks (1D-CNN) and long short-term memory (LSTM) networks. A weighted safety scoring mechanism is further proposed to aggregate window-level predictions into a temporally consistent trip-level safety measure, allowing improved sensitivity to high-risk driving events. The proposed framework is evaluated on the UAH-DriveSet dataset using leakage-free 5-fold trip-level cross-validation. Experimental results demonstrate that the 1D-CNN achieves a window-level recall of 0.91 and an AUC of 0.94, while the LSTM attains a recall of 0.89 and an AUC of 0.93. In addition, cross-domain validation on a synthetic CAN-like dataset confirms the robustness and generalization capability of the proposed approach. The obtained results indicate that the proposed framework provides an effective and interpretable solution for real-time driving safety assessment using vehicle sensor data.