Heterogeneous Graph-based Trajectory Prediction using Local Map Context and Social Interactions (2311.18553v1)

Abstract: Precisely predicting the future trajectories of surrounding traffic participants is a crucial but challenging problem in autonomous driving, due to complex interactions between traffic agents, map context and traffic rules. Vector-based approaches have recently shown to achieve among the best performances on trajectory prediction benchmarks. These methods model simple interactions between traffic agents but don't distinguish between relation-type and attributes like their distance along the road. Furthermore, they represent lanes only by sequences of vectors representing center lines and ignore context information like lane dividers and other road elements. We present a novel approach for vector-based trajectory prediction that addresses these shortcomings by leveraging three crucial sources of information: First, we model interactions between traffic agents by a semantic scene graph, that accounts for the nature and important features of their relation. Second, we extract agent-centric image-based map features to model the local map context. Finally, we generate anchor paths to enforce the policy in multi-modal prediction to permitted trajectories only. Each of these three enhancements shows advantages over the baseline model HoliGraph.

• The paper introduces a novel graph-based model that integrates rich local map context with social interactions for predicting vehicle trajectories.
• It employs an autoencoder to compress high-dimensional map features, efficiently incorporating lane markings and road boundaries.
• Evaluation on nuScenes demonstrates improved prediction accuracy with anchor paths, ensuring trajectories remain realistic and legally compliant.

Introduction

Autonomous vehicles (AVs) are poised to revolutionize transportation by enhancing safety and environmental sustainability. However, this revolution hinges on the AVs' ability to predict the movements of surrounding traffic participants effectively. Existing prediction models tend to focus on simple interactions among traffic agents but often overlook the nuanced relationships between those agents, the driving context, and traffic rules.

The Challenge of Comprehensive Traffic Modeling

Current prediction models typically use a vector-based approach, which encapsulates agents and the road network as sequences of vectors. While this method can handle simple interactions, it fails to capture intricate details like the type of relationship between agents and contextual road information, such as lane dividers. These limitations can lead to predictions that may not respect actual road topology or feasible driving behaviors.

A Novel Approach

To address these issues, a new method is presented that leverages three key enhancements:

1. Semantic Scene Graph (SSG): This graph emphasizes the connectivity and relationships between traffic participants, accounting for both the nature of their interactions and the distance along the path they share.
2. Local Map Context (MA): The method integrates agent-centric, image-based map features that provide additional details about the local driving scene, such as lane markings and road boundaries. This is achieved by using an autoencoder that compresses high-dimensional maps into a more manageable and informative latent space.
3. Anchor Paths: The model generates anchor paths that define the road segments and lanes a vehicle is allowed to occupy. These paths serve as constraints within the predictions, ensuring that proposed trajectories remain realistic and within the legal driving space.

Evaluation and Findings

The evaluation was carried out on the nuScenes dataset, which contains a diverse set of real-world traffic scenarios. The results demonstrated substantial improvements over the baseline model, particularly when anchor paths and local map contextual information were utilized. Interestingly, while the Semantic Scene Graph reduced the complexity by limiting necessary graph edges, it did not improve prediction performance. This might stem from the prevalence of traffic light-controlled intersections within the dataset that the SSG could not factor in effectively.

Conclusion

The innovative additions of the Semantic Scene Graph, local map context, and anchor paths have proven to be potent tools for enhancing the capabilities of trajectory prediction in AVs. While the SSG simplifies the model without sacrificing performance, the other two enhancements significantly refine prediction accuracy. These improvements in handling detailed information about traffic participants and road topology are poised to make autonomous driving safer and more reliable.