Matching Explanation Detail to Scene Complexity: Studying Situational Awareness-Specific AI Feedback in Pedestrian Encounter Driving Scenarios

Paper Title

Matching Explanation Detail to Scene Complexity: Studying Situational Awareness-Specific Feedback in Pedestrian Encounter Driving Scenarios

Publication Info

• Topic area: Human-Computer Interaction (HCI) in automated vehicle interfaces
• Keywords: Automated vehicles, situational awareness, pedestrian encounters, scene complexity, intent uncertainty, in-vehicle feedback, user preferences, cognitive load, driving decision entropy, human-centered design

Background and Problem

• Problem / challenge: Existing in-vehicle feedback systems often fail to adapt the level of detail to the perceived complexity of driving scenarios, leading to information overload in simple situations or insufficient detail in complex ones.
• Significance: Adapting feedback detail to scene complexity can improve user understanding, trust, and cognitive load management in automated vehicles, particularly in safety-critical pedestrian encounters.
• Motivation and related work: Prior studies have explored feedback presentation styles, timing, and content, but have not adequately addressed dynamic adaptation of feedback detail based on scene complexity. This paper builds on situational awareness (SA) theory and introduces intent uncertainty as a proxy for scene complexity in pedestrian interactions.

Solution

• Proposed approach: Introduce Perceived Scene Complexity (PSC) based on intent uncertainty in pedestrian encounters and evaluate user preferences for feedback detail using manually crafted situational awareness-specific feedback concepts.
• Novelty:
  1. Introduces intent uncertainty as a human-centered measure of scene complexity validated through driving decision diversity.
  2. Demonstrates empirically that user preferences for feedback detail scale with scene complexity, with richer feedback preferred in high-complexity scenarios.
• Procedure and key techniques:
  • Experiment 1: Measure intent uncertainty and its correlation with driving decision entropy and perceived agreement gap using dashcam videos labeled with pedestrian crossing intent.
  • Experiment 2: Evaluate user preferences for three feedback concepts (perception-only, perception + comprehension, perception + comprehension + projection) across varying levels of scene complexity using a mock-up in-vehicle interface.

Results

• Concrete findings:
  • Higher intent uncertainty correlates with increased driving decision entropy (low: 0.82, medium: 0.98, high: 1.07) and perceived agreement gap (low: 23.02%, medium: 26.49%, high: 30.65%).
  • Users consistently preferred the most detailed feedback (Concept C) across all complexity levels, with Concept A rated progressively lower as uncertainty increased.
  • Feedback with richer situational awareness detail reduced mental demand, effort, and frustration while improving perceived performance (NASA-TLX results).
• Advantage over baselines:
  • Concept C (perception + comprehension + projection) outperformed simpler feedback designs (Concept A and B) in usefulness, trust, detail sufficiency, and presentation preference across all complexity levels.
• Experiments / evaluation:
  • Experiment 1: 68 participants analyzed 24 dashcam videos to validate intent uncertainty as a proxy for scene complexity.
  • Experiment 2: 60 participants evaluated feedback concepts across 12 videos representing low, medium, and high complexity; cognitive load assessed using NASA-TLX.
• Limitations and future work:
  • Limited ecological validity due to passive video observation; future studies should use immersive environments like VR or driving simulators.
  • Findings based on manually crafted feedback; future work should account for AI model uncertainty and errors.
  • Focused exclusively on pedestrian encounters; generalization to other driving scenarios is needed.
  • Participant diversity was limited; broader demographic studies are recommended.

Summary

This paper introduces intent uncertainty as a proxy for scene complexity in pedestrian encounters and demonstrates that user preferences for in-vehicle feedback detail scale with complexity. Experiment 1 validates the correlation between intent uncertainty and driving decision diversity, while Experiment 2 shows that richer situational awareness-specific feedback improves understanding and reduces cognitive load in complex scenarios. These findings provide actionable insights for designing adaptive, human-centered automated vehicle interfaces that balance transparency and cognitive demand.