Towards Human-Like Trajectory Prediction for Autonomous Driving: A Behavior-Centric Approach

TL;DR

This work presents HiT, a human-like trajectory predictor for autonomous driving that combines a behavior-aware module, an interaction-aware module, and a multimodal decoder. It introduces dynamic geometric graphs and a novel dynamic degree centrality to capture real-time interactions and driver behavior, augmented by a fuzzy inference system and q-ROFWEBM defuzzification to produce robust aggressiveness scores. Behavior is encoded via a hypergraph to model higher-order relations, while the interaction module uses polar pooling and attention to capture spatio-temporal dynamics, culminating in a multimodal Gaussian mixture output over future ego trajectories. HiT demonstrates superior accuracy and efficiency across diverse real-world datasets, including improved data efficiency (good performance with as little as 25% of the data) and fast inference, highlighting its potential for real-time, safe autonomous driving in complex environments.

Abstract

Predicting the trajectories of vehicles is crucial for the development of autonomous driving (AD) systems, particularly in complex and dynamic traffic environments. In this study, we introduce HiT (Human-like Trajectory Prediction), a novel model designed to enhance trajectory prediction by incorporating behavior-aware modules and dynamic centrality measures. Unlike traditional methods that primarily rely on static graph structures, HiT leverages a dynamic framework that accounts for both direct and indirect interactions among traffic participants. This allows the model to capture the subtle yet significant influences of surrounding vehicles, enabling more accurate and human-like predictions. To evaluate HiT's performance, we conducted extensive experiments using diverse and challenging real-world datasets, including NGSIM, HighD, RounD, ApolloScape, and MoCAD++. The results demonstrate that HiT consistently outperforms other top models across multiple metrics, particularly excelling in scenarios involving aggressive driving behaviors. This research presents a significant step forward in trajectory prediction, offering a more reliable and interpretable approach for enhancing the safety and efficiency of fully autonomous driving systems.

Figures (6)

• Figure 1: Comparison of our proposed behavior-aware pooling mechanism using polar coordinates (left) versus the traditional fixed-size grid pooling mechanism using Cartesian coordinates (right). In the polar coordinate system, the ego vehicle (yellow) is centered at the origin, with surrounding vehicles (blue) mapped based on their relative radial distance and angle. This allows for a flexible, position-aware representation that better captures dynamic interactions, particularly in complex driving scenarios. The classical Cartesian grid-based approach (right) divides the environment into fixed-size cells, which may not accurately reflect the spatial relationships and dynamic behaviors in irregular or unstructured environments.
• Figure 2: Architecture of the proposed HiT model. The framework consists of three key components: the behavior-aware module, the interaction-aware module, and the multimodal decoder. The behavior-aware module leverages refined centrality measures to dynamically capture the evolving driving behavior of traffic agents. A novel behavior-aware criterion evaluates the intensity and patterns of these behaviors. Using these criteria, a fuzzy inference system quantifies driver aggressiveness, whereas a hypergraph neural network within the behavior encoder integrates and refines behavior features to generate behavior vectors $\mathbf{O}_{\text{beh}}$. Moreover, the interaction-aware module adopts a lightweight design with progressive positional encoding to generate spatio-temporal interaction vectors $\mathbf{O}_{\text{int}}$. Finally, the multimodal decoder fuses these vectors to predict multimodal trajectories $\bm{Y}_{0}^{t:t+t{f}}$.
• Figure 3: Membership functions of (a) BIE, (b) BFE, and (c) Aggressive score. "Low", "Medium", and "High" denote the Low-level, Medium-level, and High-level categories for the fuzzy sets, respectively.
• Figure 4: Interpretability analysis of our proposed model HiT's trajectory predictions in complex traffic scenarios. The figure presents HiT’s predictive performance in three challenging driving scenarios: (a) a right-lane change maneuver with cooperative adjacent vehicles, (b) a right-lane change maneuver where an adjacent vehicle exhibits aggressive behavior, and (c) a lane-keeping scenario with varied interactions from surrounding vehicles. The ego vehicle is highlighted in red, whereas its surrounding agents are marked in blue. The heatmaps on the right illustrate the Behavior Intensity Estimate (BIE) and Behavior Fluctuation Estimate (BFE) for each scenario, showing how HiT evaluates the influence of surrounding vehicles based on their proximity and driving behavior. The results demonstrate HiT's ability to discern subtle social interactions and predict trajectories that closely align with the actual driving behavior, outperforming other models that do not incorporate behavior-awareness.
• Figure 5: Multimodal maneuver prediction framework with corresponding probability outputs. Regarding maneuver uncertainty, we categorize the possible maneuvers based on the characteristics of the driver's actions and enable the model to predict based on these potential maneuvers. These maneuvers include speed-related maneuvers: accelerating, braking, and maintaining speed, and position-related maneuvers: left lane change, right lane change, and lane keeping.

Theorems & Definitions (10)

• Definition 1
• Definition 2
• Theorem 1
• Definition 3
• Definition 4
• Theorem 2
• Definition 5
• Definition 6
• Theorem 3
• Theorem 4