Driving Behavior Pattern Recognition using Transformer Attention Mechanisms in Vehicle Control Systems

Driving behavior pattern recognition plays a vital role in enhancing vehicle control systems, especially for applications in safety, autonomy, and intelligent transport. The integration of transformer attention mechanisms presents a promising avenue for understanding complex, time-dependent driver behaviors. Existing methods, including CNNs and classical neural networks, often struggle with limited real-world data availability, lack of generalizability, and poor interpretability in heavy-duty vehicles and naturalistic settings. To address these challenges, this paper proposes a novel framework, Detecting Driving Behaviors using Transformer with Self-Attention Mechanism (DDB-T-SAM). This method captures temporal dependencies and contextual correlations in multimodal driver input data (e.g., GPS, steering, acceleration, and facial expressions), enabling robust recognition of behaviors such as lane shifting, turning, and drowsiness. The proposed DDB-T-SAM model leverages self-attention layers to dynamically focus on critical behavioral cues across time sequences, outperforming traditional architectures in adaptability and accuracy. It is also scalable across different vehicle types and environments. Experimental evaluation on the CARLA driver behaviour dataset shows that DDB-T-SAM outperforms baseline models (NNM, DBA, DLT) by achieving the lowest average acceleration error of 0.26 m/s<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, highest steering sharpness detection accuracy of 94.0%, driving style classification accuracy of up to 92.5%, and behavior transition accuracy of 90.3%.

Ҳали таржима қилинмаган