2024-12-10 ジョージア大学 (UGA)
Self-driving cars have to accurately anticipate the movements of surrounding traffic. (Getty Images)
＜関連情報＞
- https://news.uga.edu/making-self-driving-cars-safer/
- https://news.uga.edu/making-self-driving-cars-safer/
コネクテッドカーと自動運転車の安全性を考慮したニューラルネットワーク Safety aware neural network for connected and automated vehicle operations
Handong Yao, Xiaopeng Li, Qianwen Li, Chenyang Yu
Transportation Research Part E: Logistics and Transportation Review Available online: 27 September 2024
DOI:https://doi.org/10.1016/j.tre.2024.103780
Highlights
- Introduce a novel safety-aware neural network framework for CAV operations.
- Create recurrent neural cells embedded with car-following dynamics.
- Employ time geography theory principles for safety regulation.
- Yield robust solutions with the sequential unconstrained minimization technique.
- Achieve substantial enhancements in safety and energy efficiency.
Abstract
Contemporary research in connected and automated vehicle (CAV) operations typically segregates trajectory prediction from planning in two segregated models. Trajectory prediction narrowly focuses on reducing prediction errors, disregarding the implications for subsequent planning. As a result, CAVs adhering to trajectories planned based on such predictions may collide with surrounding traffic. To mitigate such vulnerabilities, this study introduces a holistic safety-aware neural network (SANN) framework, representing a paradigm shift by integrating trajectory prediction and planning into a cohesive model. The SANN architecture incorporates prediction and planning layers, leveraging existing neural networks for prediction and introducing novel recurrent neural cells embedded with car-following dynamics for planning. The prediction layers are regulated by the CAV trajectory planning performance including safety, mobility, and energy efficiency. A key innovation of the SANN lies in its approach to safety regulation, which is based on actual, rather than forecasted, traffic movements. By applying time geography theory, it assesses CAV motion feasibility, setting limits on speed and acceleration for safety in line with actual traffic patterns. This feasibility analysis results are integrated into the neural loss function as a penalty factor, steering the optimization process towards safer CAV operations. The efficacy of the SANN is enhanced by employing the sequential unconstrained minimization technique, which enables the fine-tuning of penalty weights, thereby producing more robust solutions. Empirical evaluations, comparing the holistic SANN with conventional segregated models, demonstrate its superior performance. The SANN achieves substantial enhancements in safety and energy efficiency, with only a marginal compromise on mobility. This success underscores the significance of integrating machine learning with domain knowledge (operations research and traffic flow theory) for safer and more environmentally friendly CAV operations.
