A multibranch risk assessment model for multi-lane highway work zones based on risk fields and graph attention network

This paper addresses the challenge of real-time safety assessment in multi-lane highway work zones by introducing two key innovations. First, we formulate a Comprehensive Collision Risk Index (CCRI) that synergistically combines Equivalent Time-to-Collision (ETTC), Time-to-Dangerous-Collision (TDTC) and normalized collision energy into a single, physically interpretable metric. Second, we map discrete conflict events into a continuous risk field via an adaptive, bandwidth-weighted Kernel Density Estimation (KDE), enabling dynamic, spatially resolved delineation of low/medium/high risk zones. Building upon these indicators, we design a multibranch deep fusion network that fuses temporal cues, graph-based interactions, risk-field patches, and global traffic context via an attention-based fusion module. Evaluated on UAV-derived vehicle trajectories from the Kunshan segment of the Beijing–Shanghai Expressway, the proposed framework achieves 90.6% overall accuracy, a macro-AUC of 0.97, and a high-risk recall of 79.4%, representing improvements of 5%–12% points over single-modality and early-fusion baselines. Ablation studies confirm that CCRI and the risk-field branch contribute most critically to performance gains. The approach supports real-time warnings, variable speed limits, and lane-closure planning in complex work-zone environments.