Zhu Dianchen, Fan Zheyan, Ma Wei, Zhang Xuxin, Chan Ho-Yin, Zhao Mingming
School of Automobile and Traffic Engineering, Hefei University of Technology, Hefei, Anhui, China.
School of Civil and Transportation Engineering, South China University of Technology, Guangzhou, Guangdong, PR China.
Traffic Inj Prev. 2025;26(5):567-576. doi: 10.1080/15389588.2024.2443008. Epub 2025 Feb 18.
Intersections represent critical points where conflicts between vulnerable road users (VRUs) and vehicles often occur, posing significant safety challenges globally. Despite efforts to mitigate heterogeneous traffic individual conflict, the accurate trajectory prediction of VRU-vehicle interactions remains elusive due to asymmetric information, unequal risks, and uncertainties in decision-making behaviors. Most existing trajectory prediction models predominantly focus on either VRUs or vehicles, neglecting the complex mechanisms of interactions between heterogeneous traffic agents. This article proposes a trajectory prediction model that incorporates spatial-temporal characteristics and security risk awareness as an alternative approach to these challenges.
The proposed spatial-temporal risk network (STRN) combines the awareness of time, space, and quantified safety risk to improve the performance of this model. First, the safety potential field theory is used to quantify and label the risks in the VRU-vehicle interaction scenario, and the effect of conflict risk on the agents' motion trajectory is considered. Second, the spatial constraint features of agent movement are extracted from the spatial dimension. Third, the change characteristics of the agent's motion trajectory over time are extracted from time dimension.
The experimental results show that the model can effectively identify the motion trajectories under the complex interaction between VRUs and vehicles. The ablation experiments confirm that the integration of spatial-temporal risk dimensions positively impacts the accuracy of interaction prediction. The model shows great robustness in different scenario transferability tests.
This study revisits the challenge of predicting interaction trajectories of heterogeneous objects in unsignalized intersection scenarios, a topic that has not been extensively explored. The proposed STRN model, with its performance and robustness, provides a new scheme for improving the level of traffic safety and promoting intelligent autonomous vehicle decision system.
交叉路口是弱势道路使用者（VRU）与车辆之间冲突频发的关键点，在全球范围内构成了重大安全挑战。尽管人们努力缓解异质交通中的个体冲突，但由于信息不对称、风险不均以及决策行为的不确定性，VRU与车辆交互的精确轨迹预测仍然难以实现。现有的大多数轨迹预测模型主要侧重于VRU或车辆，忽略了异质交通主体之间复杂的交互机制。本文提出了一种结合时空特征和安全风险意识的轨迹预测模型，作为应对这些挑战的替代方法。
所提出的时空风险网络（STRN）结合了时间、空间和量化安全风险的意识，以提高该模型的性能。首先，利用安全势场理论对VRU与车辆交互场景中的风险进行量化和标记，并考虑冲突风险对主体运动轨迹的影响。其次，从空间维度提取主体运动的空间约束特征。第三，从时间维度提取主体运动轨迹随时间的变化特征。
实验结果表明，该模型能够有效识别VRU与车辆复杂交互下的运动轨迹。消融实验证实，时空风险维度的整合对交互预测的准确性有积极影响。该模型在不同场景可转移性测试中表现出很强的鲁棒性。
本研究重新审视了无信号交叉路口场景中异质物体交互轨迹预测的挑战，这一主题尚未得到广泛探索。所提出的STRN模型具有良好的性能和鲁棒性，为提高交通安全水平和推动智能自动驾驶车辆决策系统提供了一种新方案。
