Cell-Trans

Abstract

Intelligent vehicle cyberphysical systems can integrate real-time traffic scene perception with built-in high-definition maps to construct digital twins of real-world signalized intersections. Based on digital twins, this paper presents a traffic prediction method named cell transformer (cell-trans), comprising vehicle-, cell-, and road-level encoders and a decoder. The vehicle-level encoder first converts vehicle features into vehicle encodings, which the cell-level encoder then fuses with lane segment features to generate cell encodings. Next, the road-level encoder treats the connectivity between lane segments and the phase information at signalized intersections as a dynamic directed graph, extracting spatial-temporal evolution patterns to improve traffic predictions. The cell-trans is compared with baseline models on pNEUMA and CitySim data sets, and the performance comparison validates its optimal predictive accuracy. Moreover, the outstanding performance of the cell-trans is confirmed by ablation study, parameter analysis, and computational efficiency analysis. Finally, this paper develops a cell-trans-based motion planner for autonomous vehicles (AV) in a joint simulation platform combined CARLA and SUMO to indicate its contributions to AVs.