Dynamic Aware: Adaptive Multi-Mode Out-of-Distribution Detection for Trajectory Prediction in Autonomous Vehicles

Abstract

Trajectory prediction is central to the safe and seamless operation of autonomous vehicles (AVs). In deployment, however, prediction models inevitably face distribution shifts between training data and real-world conditions, where rare or underrepresented traffic scenarios induce out-of-distribution (OOD) cases. While most prior OOD detection research in AVs has concentrated on computer vision tasks such as object detection and segmentation, trajectory-level OOD detection remains largely underexplored. A recent study formulated this problem as a quickest change detection (QCD) task, providing formal guarantees on the trade-off between detection delay and false alarms [1]. Building on this foundation, we propose a new framework that introduces adaptive mechanisms to achieve robust detection in complex driving environments. Empirical analysis across multiple real-world datasets reveals that prediction errors -- even on in-distribution samples -- exhibit mode-dependent distributions that evolve over time with dataset-specific dynamics. By explicitly modeling these error modes, our method achieves substantial improvements in both detection delay and false alarm rates. Comprehensive experiments on established trajectory prediction benchmarks show that our framework significantly outperforms prior UQ- and vision-based OOD approaches in both accuracy and computational efficiency, offering a practical path toward reliable, driving-aware autonomy.

Figures (7)

• Figure 1: Illustration of our adaptive multi-mode OOD detection framework for trajectory prediction in AVs. The framework monitors prediction errors in real-time, dynamically identifies the active error regime (low-error or high-error mode), and adapts detection thresholds accordingly.
• Figure 2: Multi-modal distributions of prediction errors. The $x$-axis denotes prediction error (log-scale ADE; see Section \ref{['sec:metrics']}), and the $y$-axis denotes probability density. Plots were averaged over $500$ independent runs, the empirical distribution consistently exhibits a bi-modal structure, well approximated by a two-component GMM. The lower-error component is defined as the low-risk mode, while the higher-error component represents the high-risk mode.
• Figure 3: Mode dynamics across different driving scenes (FQA model). Datasets: ApolloScape (left), NGSIM (center), nuScenes (right). The $x$-axis represents the time frame indices; the $y$-axis is error magnitude (log scale). The modes were inferred by the classic MAP estimation via the GMM posterior (blue = low-risk, orange = high-risk) hajek2015random.
• Figure 4: Observable scene-level information provides unreliable cues for error modes and their transitions. The figure illustrates (i) agent counts over time, (ii) maximum and average speeds over time, and (iii) corresponding map snapshots of the driving scenes. The pink dashed lines indicate where a mode switch occurs.
• Figure 5: Three evaluation datasets: ApolloScape (dense urban interactions), NGSIM (freeway driving), and nuScenes (multimodal urban scenes).