LaTAS-F: Locality-aware transformer architecture search with multi-source fusion for driver continuous braking intention inference

Abstract

Precise inference of driver braking intention is highly correlated with traffic safety, energy consumption, and driving comfort of electrified vehicles (EVs). Until recently, gratifying results have been achieved in accurately inferring braking intention as discrete states, such as classify brake or no brake, and judge normal or emergency brake. However, accurately and quantitively predicting driver brake intensity is significant for enhancing vehicle safety, its related research is especially rarely seen. To mitigate this deficiency, a novel framework, locality-aware transformer architecture search with multi-source fusion (LaTAS-F), is put forward for continuous braking intention (CBI) inference in this paper. To consider the comprehensively complicated 'driver-vehicle' interactions, data composed by driver physical state and vehicle state are collected with multi-source sensors. A kernel principal component analysis (KPCA) method is proposed to fuse multi-source data while removing redundancies. In addition, an improved transformer with enhanced locality-aware attention mechanism (ELAM) is elaborately designed to alleviate the incapability of self-attention mechanism (SAM) in canonical transformer when capturing local context. In particular, due to the inefficiency of designing the transformer entirely by hand, a tree-structured parzen estimator (TPE) strategy is implemented to automatically design the most effective and practical model architecture. The performance of LaTAS-F framework is evaluated in various of historical and predicted horizons. Data is collected from 21 subjects, and experimental results demonstrate that our proposed framework can accurately predict CBI of 200 ms in the future, outperforming baselines with RMSE of 0.428Mpa and R2 of 0.963.