Autonomous driving systems (ADS) increasingly rely on vision-based perception and learning-based decision modules, yet real-world safety depends on whether these systems remain stable as the ego vehicle continuously changes distance and viewpoint relative to roadway objects. Current evaluation practices often emphasize single images or limited viewpoints, which can mask trajectory-dependent failure modes that emerge during real driving under changing illumination, partial occlusion, and motion blur.
This project studies the limitations of existing ADS through a measurement-driven, multi-view robustness evaluation framework designed to produce actionable engineering insights and evidence-based inputs for transportation safety policy. Building on a differentiable, view-consistent scene representation (3D Gaussian Splatting with view-dependent appearance modeling), the team will generate controlled, physically plausible appearance variations across realistic approach trajectories and use them as a diagnostic tool to quantify perception instability and downstream planning sensitivity.
The project will deliver a reproducible set of safety-relevant scenarios and ego-vehicle approach trajectories representing how a vehicle observes the same object over time; a controllable multi-view rendering and perturbation engine built on 3D Gaussian Splatting that synthesizes viewpoint-consistent observations under bounded, physically plausible appearance variations; and a multi-view robustness evaluation methodology and benchmark protocol using trajectory-based sampling. It will produce quantitative robustness indicators summarizing perception stability and planning sensitivity, a structured taxonomy of observed failure modes, and a reproducible reporting package of metrics definitions, evaluation scripts, and documentation templates. Validation will be performed on representative research ADS models, with black-box evaluation of commercial systems where feasible and safe.
Universities Involved
University of Maryland, College Park
Principal Investigators
Xianfeng Yang
Expected Research Outcomes & Impacts
The project advances both ADS engineering practice and traffic safety by moving robustness discussions from single-image or single-view tests to trajectory-consistent, safety-relevant evaluations that are closer to real driving operations. By identifying repeatable failure patterns and their triggering conditions, it will generate practical insights for improving ADS design, including better training coverage for viewpoint diversity, stronger temporal consistency mechanisms, uncertainty-aware decision making, and improved sensor fusion.
The benchmark-style outputs reduce duplicated effort across organizations and enable more transparent communication of limitations, improving the credibility of safety claims. For transportation agencies, the evidence-based artifacts can inform policy and regulation, strengthen how operational design domain readiness and validation evidence are discussed, and highlight roadway conditions, such as approach-angle sensitivity, distance-dependent recognition instability, and lighting-driven degradation, that deserve explicit attention in testing guidance.
Subject Areas
Autonomous Vehicles, Computer Vision, Safety Evaluation, Machine Learning, Transportation Policy
