Human-Like Attention Allocation for AV Car-Following Control in Lane-Changing Interactions

Lane-changing interactions pose complex car-following (CF) challenges for automated vehicles (AVs), as AVs acting as following vehicles must adaptively respond to both the lane-changing vehicle and their current leader. Existing CF models for lane-changing scenarios primarily rely on fixed interaction mechanisms or predefined behavioural rules, limiting their adaptability to complex and dynamic lane-changing interactions. To address this limitation, this paper proposes the Human-Like Attention Allocation-based Car-Following (HLA-CF) framework for AVs in lane-changing interaction scenarios. Specifically, a hybrid framework is developed by integrating a physics-based CF model with an attention-based neural network, in which the neural network is embedded into the physical model to learn a human-like attention allocation mechanism. The physical model and neural network are encoded into a unified computational graph, enabling end-to-end joint learning of model parameters. Experiments on two real-world trajectory datasets, CitySim and highD, show that HLA-CF achieves lower trajectory prediction errors than the best-performing baselines, with error reductions across speed and position MAE/RMSE metrics ranging from 10.83% to 16.08% on CitySim and from 6.48% to 7.25% on highD. Theoretical analysis reveals that state-dependent attention allocation plays an important role in system stability of vehicles: disturbances can affect stability through the attention channel, while the gradients of attention weights with respect to different traffic-state features exert distinct nonlinear effects on stability. Simulation results further demonstrate the effectiveness of the human-like attention allocation mechanism in improving safety and comfort and in mitigating traffic oscillations. The results also show that HLA-CF exhibits strong adaptability and robustness under varying traffic conditions. These findings highlight the important role of attention allocation mechanisms in shaping car-following behaviour and provide theoretical insights into AV car-following control in complex lane-change interaction scenarios.