Infrastructure-enabled centralized trajectory planning for lane sorting in mixed-autonomy urban traffic bottlenecks

In urban traffic bottlenecks, vehicles are frequently compelled to execute mandatory lane changes (MLCs) to align with their intended destinations. This process is referred to as lane sorting, which is ubiquitous in scenarios such as intersection approaches and highway weaving sections. However, uncoordinated MLC maneuvers among multiple vehicles can mutually influence one another, thereby inducing severe inter-vehicle conflicts and traffic oscillations. While individual vehicle intelligence is limited by localized decision-making, vehicle-infrastructure cooperation offers a promising solution to achieve the global optimum. In this article, we propose an infrastructure-enabled centralized multi-vehicle trajectory planning framework for lane sorting in mixed-autonomy traffic bottlenecks, composed of connected and autonomous vehicles (CAVs) and human-driven vehicles (HDVs). The complex car-following relationships are explicitly modeled for scenarios where multiple vehicles execute simultaneous lane changes across multiple lanes. The two-dimensional interaction models for CAVs and HDVs are presented differently, enabling the modeling of intricate longitudinal and lateral interactions among heterogeneous vehicles. Subsequently, a mixed-integer quadratic programming (MIQP) model is formulated for multi-vehicle trajectory planning. To tackle the computational issue of the MIQP model, a heuristic multi-stage algorithm (HMA) is proposed to obtain near-optimal solutions. Complete reference trajectories with a higher temporal resolution are generated using quintic polynomial curves. Experimental results demonstrate that the proposed infrastructure-enabled centralized planning methods effectively mitigate the speed reductions and traffic disturbances induced by MLCs during the lane-sorting process, thereby enhancing overall traffic efficiency relative to the baseline approaches. Moreover, the HMA reduces the computation time to an acceptable level in high-demand cases, with the solution quality deviating only slightly from the global optimum.