Building Real-time Awareness of Out-of-distribution in Trajectory Prediction for Autonomous Vehicles

TL;DR

This work tackles real-time OOD in autonomous-vehicle trajectory prediction by recasting reliability monitoring as a quickest change-point detection problem. It leverages CUSUM, supported by Gaussian-mixture models for pre- and post-change errors, and explores complete, partial, and unknown-knowledge regimes, including a robust shift-based variant for worst-case scenarios. Through experiments on ApolloScape, NGSIM, and NuScenes with two state-of-the-art predictors, the authors demonstrate that CUSUM-based detectors—especially the full-knowledge variant (CUSUM Mix)—achieve the fastest, most reliable detection with controlled false alarms, even for deceptive OOD scenes. The results highlight the practical potential of online OOD awareness to enable timely safety interventions in AV systems, with implications for safer handoffs and adaptive control. The approach combines theoretical guarantees from change-point detection with domain-specific modeling of trajectory prediction errors to deliver a scalable, real-time safety mechanism.

Abstract

Accurate trajectory prediction is essential for the safe operation of autonomous vehicles in real-world environments. Even well-trained machine learning models may produce unreliable predictions due to discrepancies between training data and real-world conditions encountered during inference. In particular, the training dataset tends to overrepresent common scenes (e.g., straight lanes) while underrepresenting less frequent ones (e.g., traffic circles). In addition, it often overlooks unpredictable real-world events such as sudden braking or falling objects. To ensure safety, it is critical to detect in real-time when a model's predictions become unreliable. Leveraging the intuition that in-distribution (ID) scenes exhibit error patterns similar to training data, while out-of-distribution (OOD) scenes do not, we introduce a principled, real-time approach for OOD detection by framing it as a change-point detection problem. We address the challenging settings where the OOD scenes are deceptive, meaning that they are not easily detectable by human intuitions. Our lightweight solutions can handle the occurrence of OOD at any time during trajectory prediction inference. Experimental results on multiple real-world datasets using a benchmark trajectory prediction model demonstrate the effectiveness of our methods.

Figures (10)

• Figure 1: Fig. 2. Illustration of overrepresented and underrepresented scenes for the deployed ML models. Images from PTV Vissim simulation.
• Figure 2: Fig. 3. Visualization of minor perturbation applied to the history trajectory of the target vehicle. The input and output trajectories consist of $L_I = 6$ and $L_O = 6$ time frames, respectively, while $L_P = 1$ denotes the perturbed time frame within the input. Dots represent observed and predicted trajectory positions at different time frames.
• Figure 3: Fig. 4. Illustration of the performance comparison of a given ML trajectory prediction model on ID and OOD scenes. In an ID scenario (left figure), the ego vehicle (gray car), on which the ML model is implemented, can accurately predict the trajectories of neighboring vehicles. In an OOD scenario (right figure), the target vehicle slightly deviates from its usual driving path due to unexpected debris from the vehicle ahead. As a result, the ego vehicle may mispredict the target vehicle's trajectory, mistakenly assuming it will move into the same lane. In response, the ego vehicle may suddenly brake, potentially causing a rear-end collision.
• Figure 4: Fig. 5. Illustration of the mixture Gaussian distributions for pre-change (red) and post-change (blue) error distributions in trajectory prediction, assessed using the ADE metric on the ApolloScape dataset. (a) Results from the GRIP++ model. (b) Results from the FQA model. Both examples highlight the observed shift in error distributions.
• Figure 5: Fig. 6. For the actual gap $\eta = 2.5$, we set the minimum test shift parameter to $\kappa = 1$ to evaluate robustness. The robust test (blue curve) with shift = 1 achieves successful detection, validating its effectiveness. In contrast, the non-robust tests ( pink curve with shift = 5 and orange curve with shift = 10) fail to maintain detection capability, as their parameter choices violate the robustness criteria.

Theorems & Definitions (3)

• Theorem 1: veeravalli2014quickest
• definition 1
• Theorem 2