FSTAN: A Two-Stage Traffic Conflict Prediction Framework Based on Vehicle Trajectory

Microscopic traffic conflict prediction is critical to improving road safety. For autonomous driving in highway scenarios, we develop a two-stage conflict prediction framework. In the first stage, a field force-based spatiotemporal attention network field-based spatial-temporal attention network (FSTAN) for trajectory prediction is proposed, which models vehicle interaction perception as a measurement process. The FSTAN network applies spatiotemporal transformer blocks (STTBs) with multiple attention mechanism to capture complex vehicle interactions. A potential field force attention mechanism is designed to capture spatial interactions, while self-attention mechanism is applied to model temporal interactions. The field force attention mechanism imparts physical significance to spatial interactions through field risk forces, enhancing the model’s interpretability. In the second stage, a bounding box-based conflict recognition method is introduced, which takes into account the influence of vehicle geometry on conflict recognition. Experimental results highlight the superior long-term prediction capability of FSTAN compared to several state-of-the-art (SOTA) models. Visualization results demonstrate that the bounding box-based measurement method outperforms the center point-based method in both side-on and rear-end conflict detection. Center point-based measurement methods tend to underestimate collision risks and miss potential conflicts by neglecting vehicle geometry. The proposed two-stage framework lays a foundation for further development of real-world vehicle warning systems.

I. Introduction

The synergy between artificial intelligence (AI) and the internet of things (IoT) has accelerated the development of autonomous vehicle (AV) technology. Internet of vehicles (IoV) technology enables connected vehicles to communicate with edge servers, allowing them to access real-time information about their surroundings. Through sensing and communication, autonomous driving aims to enhance measurement integrity and efficiency [1], while improving traffic system safety [2]. However, recent reports of AV accidents indicate significant limitations in autonomous driving technology. Several fatal Tesla accidents have been attributed to perception and prediction errors [3], [4]. Existing statistics show that prediction errors account for over 40% of collision factors involving traditional vehicles, which align with reports on AV crashes [5]. This underscores the critical importance of traffic conflict prediction for the advancement of autonomous driving technology.