Quantifying and Modeling Driving Styles in Trajectory Forecasting

TL;DR

This work addresses the gap in trajectory forecasting by explicitly considering human driving styles as a distinct, relatively stable factor that influences edge-case outcomes. It analyzes real-world datasets to quantify driving styles using both TDBM and unsupervised kinematic clustering (KDSC), revealing underrepresentation of fringe styles and cross-dataset variations. A style-aware forecasting framework augments a base predictor with driving-style and context embeddings, showing improved generalization and performance on edge-case scenarios, especially when style information is fused early in the encoder. The findings highlight the value of incorporating driving style to improve robustness and safety in autonomous driving predictions, while acknowledging limitations in current datasets and measurement approaches and pointing toward balanced training and stronger OOD techniques for future work.

Abstract

Trajectory forecasting has become a popular deep learning task due to its relevance for scenario simulation for autonomous driving. Specifically, trajectory forecasting predicts the trajectory of a short-horizon future for specific human drivers in a particular traffic scenario. Robust and accurate future predictions can enable autonomous driving planners to optimize for low-risk and predictable outcomes for human drivers around them. Although some work has been done to model driving style in planning and personalized autonomous polices, a gap exists in explicitly modeling human driving styles for trajectory forecasting of human behavior. Human driving style is most certainly a correlating factor to decision making, especially in edge-case scenarios where risk is nontrivial, as justified by the large amount of traffic psychology literature on risky driving. So far, the current real-world datasets for trajectory forecasting lack insight on the variety of represented driving styles. While the datasets may represent real-world distributions of driving styles, we posit that fringe driving style types may also be correlated with edge-case safety scenarios. In this work, we conduct analyses on existing real-world trajectory datasets for driving and dissect these works from the lens of driving styles, which is often intangible and non-standardized.

Figures (7)

• Figure 1: Motivating Example: yellow-light scenario. In this specific but common scenario, a vehicle approaches the intersection such that the yellow light would turn red as the vehicle enters the intersection. This scenario poses a decision to the human driver: slow down to reduce jerk when stopping, or accelerate to cross the intersection before the light turns red. Given that the intention, context, and history of the driver is the same in both outcomes, we hypothesize that multi-modal outcomes here are based on driving style differences. In this example, both the average and the final difference between the two outcomes is high. Driving style is difficult to quantify and under-explored in context of trajectory forecasting for vehicles. In this paper, we perform preliminary analysis of possible quantifications of driving style and propose an approach to address style imbalances and improve performance on rarer driving styles.
• Figure 2: Distribution of TDBM Driving Styles in Existing Trajectory Forecasting Evaluation Datasets. Driving style scoring based on the Trajectory to Driver Behavior Mapping (TDBM) cheung2018_tdbm provides a feature mapping from kinematic properties of trajectories to a driving style classification validated by a user perception study. TDBM has six total driving style classes (from most to least aggressive): aggressive, reckless, threatening, careful, cautious, and timid. The driving style distribution was plotted across four different datasets: nuScenes nuscenes, Argoverse 2 Argoverse2, DISC kumar2025_disc, and Waymo Open Motion Dataset (WOMD) Ettinger_2021_ICCV_waymo. While nuScenes, Argoverse, and WOMD represent general real-world driving cases, DISC focuses on pre-crash scenarios collected in virtual reality simulated user studies. We note that there is no presence of "aggressive", "cautious", or "timid" driving in any datasets, as defined by TDBM. This may suggest that TDBM is not adequate to capture diverse driving styles in dense urban interactions and/or edge cases.
• Figure 3: Heatmap illustrating the comparison between MDSI-derived self-reported driving styles (X-axis) and TDBM-derived trajectory-based driving styles (Y-axis) within the DISC dataset. The intensity of each cell represents the frequency of participants exhibiting a given MDSI-TDBM driving style pair. Discrepancies between self-reported and observed behaviors highlight the influence of accident scenarios on driving trajectories and kinematics.
• Figure 4: Boxplots for Second-Order Kinematics by TDBM Driving Style in the nuScenes and Argoverse 2 datasets. Different TDBM driving styles have statistically significant differences in kinematics. More aggressive driving styles (reckless/threatening) typically exhibit greater speeds, accelerations, and jerks compared to timid driving. "Careful" driving style was omitted due to an inadequate number of samples for a statistically significant analysis.
• Figure 5: Distinct Driving Styles in nuScenes, Argoverse, DISC datasets, with Mean Speed Distributions for Aggressive vs. Normal KDSC Clusters. Distinct clusters, identified using kinematic features, reveal statistically significant differences in driving styles, demonstrating that embedded dataset information can differentiate aggressive (orange) and normal (blue) driving behaviors. Although these distributions differ markedly across different datasets, they also exhibit meaningful overlap in distribution between training (top) and test (bottom) data, confirming that our KDSC approach is not naively clustering based on speeds. Note: speed was not used as a direct clustering feature, instead more subtle kinematic cues (i.e. maximum absolute acceleration, variance of acceleration, variance of speed, and $\gamma$) effectively separate different driving styles.