Dynamic Risk Assessment for Autonomous Vehicles from Spatio-Temporal Probabilistic Occupancy Heatmaps

TL;DR

The paper tackles dynamic collision risk for autonomous vehicles by introducing PORA, a probabilistic occupancy risk metric derived from spatiotemporal occupancy heatmaps. PORA combines a modular heatmap generation framework (Transformer-based encoder with GAN-based decoders) with a Cox model-based dynamic risk adjustment that leverages changes in occupancy to reflect relative motion, yielding a risk score per trajectory that can guide real-time planning. Through extensive simulations on Argoverse 2 and SUMO, PORA consistently outperforms Time-to-Collision and Density Planner baselines in learning safer policies, robustness tests, and cross-domain generalization, while maintaining real-time feasibility on GPU. The modular design, data-driven calibration, and demonstrated scalability suggest PORA as a principled, uncertainty-aware tool for AV safety evaluation and controller training in diverse traffic environments.

Abstract

Accurately assessing collision risk in dynamic traffic scenarios is a crucial requirement for trajectory planning in autonomous vehicles~(AVs) and enables a comprehensive safety evaluation of automated driving systems. To that end, this paper presents a novel probabilistic occupancy risk assessment~(PORA) metric. It uses spatiotemporal heatmaps as probabilistic occupancy predictions of surrounding traffic participants and estimates the risk of a collision along an AV's planned trajectory based on potential vehicle interactions. The use of probabilistic occupancy allows PORA to account for the uncertainty in future trajectories and velocities of traffic participants in the risk estimates. The risk from potential vehicle interactions is then further adjusted through a Cox model\edit{,} which considers the relative \edit{motion} between the AV and surrounding traffic participants. We demonstrate that the proposed approach enhances the accuracy of collision risk assessment in dynamic traffic scenarios, resulting in safer vehicle controllers, and provides a robust framework for real-time decision-making in autonomous driving systems. From evaluation in Monte Carlo simulations, PORA is shown to be more effective at accurately characterizing collision risk compared to other safety surrogate measures. Keywords: Dynamic Risk Assessment, Autonomous Vehicle, Probabilistic Occupancy, Driving Safety

Figures (23)

• Figure 1: General Approach to Collision Risk Evaluation through PORA.
• Figure 2: Proposed Modular Framework for Autonomous Vehicle Prediction Tasks.
• Figure 3: Data Flow across Perception, Prediction, and Risk Assessment Modules in the PORA Framework. This diagram illustrates the modular encoder-decoder architecture and its integration into the PORA risk assessment pipeline. Perception inputs are processed by an attention-based encoder. A task-specific decoder then generates occupancy heatmaps, which are evaluated through a risk assessment module to compute the PORA metric.
• Figure 4: Illustration of the elements in the example traffic scenario and its heatmap $H_{t_k}$: traffic participants (including AV) and map. The pink dotted line represents the AV's planned trajectory in this example scenario with the planned direction of motion. The leftmost part of the figure shows a bird's-eye view of a scenario with 4 vehicles in a 3 lane road with multiple traffic participants (green) whose future motion is to be predicted by the AV (blue). The rightmost part of the figure shows the corresponding predicted occupancy probability heatmap, where red indicates higher occupancy probability and green indicates lower probability. The prediction shown in the heatmap corresponds to the immediate next timestep, but the framework also supports longer prediction horizons. The spatial grid resolution is finer in practice; a coarser resolution is used in these examples so that the grid can be clearly visualized.
• Figure 5: Attention-Based Encoder for Modular Feature Extraction in Autonomous Vehicle Prediction: In the heatmap prediction setting, the encoder processes multiple input streams, including traffic data, map information, and target trajectory, to generate a shared latent representation that captures spatial and temporal dependencies. Each input type is encoded independently through self-attention heads, such as the traffic and map heads, which extract relevant features into an environment latent matrix. A cross-attention mechanism refines this matrix by integrating information across input sources, resulting in a unified latent matrix (pool) that serves as the contextual embedding for downstream prediction tasks. During encoder training, a pre-trained decoder with fixed parameters is attached, allowing the encoder to learn representations aligned with the decoder's input requirements.