Identifying Driver Interactions via Conditional Behavior Prediction

Tolstaya, Ekaterina; Mahjourian, Reza; Downey, Carlton; Vadarajan, Balakrishnan; Sapp, Benjamin; Anguelov, Dragomir · 2021 · Crossref

DOI: 10.1109/icra48506.2021.9561967

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of modeling interactive driving scenarios, such as lane changes and merges, which are critical for autonomous vehicle (AV) planning. The authors argue that standard passive behavior prediction models fail to account for how the AV’s future actions influence other agents. To solve this, they introduce Conditional Behavior Prediction (CBP), a framework that predicts the future trajectories of other agents conditioned on a query trajectory for the ego-agent. A key contribution is the formalization of an "interactivity score" based on mutual information, which quantifies the degree of influence between agents and aids in scenario mining and agent prioritization. The proposed method utilizes a single-shot, end-to-end deep neural network that outputs Gaussian Mixture Models representing future trajectory distributions. The model takes static scene elements and dynamic agent histories as input, along with an optional conditional query trajectory for a specific agent. It is trained on a massive proprietary dataset comprising 19 million unique scenarios and 1.9 billion vehicle agents collected over 18 years. The architecture employs an encoder for road lanes and agent history, a decoder for trajectory modes, and a Graph Neural Network (GNN) for refinement. The interactivity score is derived from the Kullback-Leibler divergence between conditional and marginal distributions, estimated via importance sampling to remain computationally tractable. Experimental results demonstrate that conditional predictions improve accuracy by 10% compared to marginal predictions, as measured by weighted Average Distance Error (wADE). The model achieves state-of-the-art performance on the Argoverse benchmark, with a minADE of 0.7488, which further improves to 0.7409 when conditioned on the sensor vehicle. The study validates the interactivity score by showing a strong correlation between high mutual information scores and "surprising interactions," where an agent’s behavior significantly changes due to another’s actions. Furthermore, the score proves effective for agent prioritization; selecting agents based on interactivity scores rather than proximity yields more accurate AV trajectory predictions under computational constraints. The significance of this work lies in providing a principled, information-theoretic definition of interactivity that is independent of specific planner definitions. By enabling single-shot conditional inference, the approach offers a tractable alternative to iterative sampling methods, which are prone to compounding errors. The interactivity score serves as a practical tool for mining interactive scenarios for training and for identifying salient agents in real-time planning, thereby enhancing the safety and efficiency of autonomous driving systems in complex, multi-agent environments.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation
• telematics crash prediction

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model