On the Assessment of Sensitivity of Autonomous Vehicle Perception

Apostol Vassilev, Munawar Hasan, Edward Griffor, Honglan Jin, Pavel Piliptchak, Mahima Arora, Thoshitha Gamage

Keywords: Sensitivity Assessment, CARLA simulation, AV Perception Robustness, Deep Ensemble, Safety Quadrant

Abstract: The viability of automated driving depends on perception systems that provide accurate and reliable real-time information for safe decision-making. However, these systems must operate not only under ideal conditions but also under challenging environments such as adverse weather, occluded roadway objects, and varying vehicle speeds, which can lead to detection errors and delayed responses. This work evaluates the robustness of autonomous vehicle perception using predictive sensitivity quantification based on an ensemble of models. We propose a notional architecture for perception evaluation that integrates multiple data sources through a ROS-based interface and an extensible AI pipeline. Experiments are conducted using CARLA-generated scenarios and real vehicle data, allowing assessment of simulation limitations. Results show that diminished lighting conditions, such as fog and low sun altitude, have the greatest impact on perception performance. Adversarial occlusions further increase model sensitivity. Greater distances to roadway objects and higher vehicle speeds also degrade perception performance, resulting in reduced perception robustness.

Submission Number: 11