A Trajectory Generator for High-Density Traffic and Diverse Agent-Interaction Scenarios

TL;DR

This work tackles the long-tail problem in trajectory prediction by generating high-density, diverse driving scenarios directly on real HD maps. It introduces HiD^2, a grid-based lane representation with behavior-aware planning and Frenet-based smoothing to produce realistic, safe, and varied multi-agent trajectories. Experiments on Argoverse 1 and 2 show HiD^2 increases scenario density and interaction diversity while preserving motion realism, and augmenting real data yields improved downstream trajectory prediction in dense settings. Overall, HiD^2 provides a scalable data-generation approach that complements real data and improves the robustness of prediction models in challenging high-density environments.

Abstract

Accurate trajectory prediction is fundamental to autonomous driving, as it underpins safe motion planning and collision avoidance in complex environments. However, existing benchmark datasets suffer from a pronounced long-tail distribution problem, with most samples drawn from low-density scenarios and simple straight-driving behaviors. This underrepresentation of high-density scenarios and safety critical maneuvers such as lane changes, overtaking and turning is an obstacle to model generalization and leads to overly optimistic evaluations. To address these challenges, we propose a novel trajectory generation framework that simultaneously enhances scenarios density and enriches behavioral diversity. Specifically, our approach converts continuous road environments into a structured grid representation that supports fine-grained path planning, explicit conflict detection, and multi-agent coordination. Built upon this representation, we introduce behavior-aware generation mechanisms that combine rule-based decision triggers with Frenet-based trajectory smoothing and dynamic feasibility constraints. This design allows us to synthesize realistic high-density scenarios and rare behaviors with complex interactions that are often missing in real data. Extensive experiments on the large-scale Argoverse 1 and Argoverse 2 datasets demonstrate that our method significantly improves both agent density and behavior diversity, while preserving motion realism and scenario-level safety. Our synthetic data also benefits downstream trajectory prediction models and enhances performance in challenging high-density scenarios.

Figures (4)

• Figure 1: Comparison of different trajectory generation methods. (A) Simulator generation: Trajectories are generated through manual operation in a simulator, which can synthesize diverse behaviors, but lacks the constraints of a real map. (B) Training-based methods: Rely on raw datasets and models, making it difficult to generate rare behaviors in the tail. (C) Our HiD$^{2}$ method: Leveraging real maps and agent information, it generates high-density scenarios and diverse rare behaviors, effectively alleviating the long-tail distribution problem without requiring extensive manual effort.
• Figure 2: Visualization of generated trajectories under progressively complex driving scenarios. (A) Original trajectories in the baseline scenario. (B) Scenario with increased vehicle density, where the framework maintains robust trajectory generation despite tighter spacing and higher interaction frequency among vehicles. (C) Trajectory generation with lane-changing behavior, demonstrating the ability to adapt to dynamic intentions, negotiate surrounding traffic, and ensure collision-free maneuvering. (D) Trajectory generation with lane-changing and overtaking behaviors, highlighting the ability to implement competitive driving strategies and generate realistic multi-agent interactions in complex environments.
• Figure 3: Comparison of dataset distributions before (pink) and after (blue) using HiD$^{2}$ for data generation. Left: distribution of scenarios across different agent density levels (increase in high-density cases). Right: distribution of scenarios with different driving behaviors (enriches the occurrence of complex interactions).
• Figure 4: Effect of adding different numbers of agents into the scenarios. From left to right are LO, LA, JE, SCR, and ORR.