1. Abstract

In the rapidly advancing field of autonomous driving, precise prediction of vehicle trajectories is paramount for ensuring safety and efficiency. Our research introduces the Safety-Aware Trajectory Prediction (SA-TP) model, a pioneering solution in trajectory prediction. This model distinctively integrates the Responsibility-Sensitive Safety (RSS) metric to encapsulate the nuanced safety attributes of traffic agents, a first in this domain. A pivotal contribution of our model is the uncertainty-aware graph attention network, which adopts a variable-learnable noise approach in graph attention network, and effectively simulates the inherent uncertainties within traffic environments, significantly enhancing SA-TP's adaptability to diverse and challenging traffic conditions. We conducted extensive testing across the Next Generation Simulation (NGSIM), Highway Drone (HighD), and the Macao Connected Autonomous Driving (MoCAD) datasets to validate SA-TP's efficacy. The results demonstrate that SA-TP achieves state-of-the-art accuracy in trajectory prediction while maintaining a relatively lightweight architecture and fast inference speed, which significantly improves the safety of autonomous driving.

2. Methodology Overview

Proposed trajectory prediction model that considers safety and uncertainty.

The detailed structure of the developed Uncertainty-aware graph attention network.

3. Results

Qualitative results are provided below.

These are visualization results on NGSIM and INTERACTION datasets.

4. Contact

If you have any questions, feel free to contact Hanlin Kong (hanlinkong@foxmail.com).