On the Assessment of Sensitivity of Autonomous Vehicle Perception

TL;DR

This work investigates the sensitivity and robustness of autonomous vehicle perception by quantifying predictive uncertainty with a deep ensemble of five state-of-the-art models on stop-sign detection under adversarial and adverse environmental conditions. It introduces a formal stopping-distance framework and a Safety Quadrant, defined by $sd$, $d_S$, $_S$, and a threshold $ heta_S$, to evaluate when perception guarantees safe braking, using ensemble mean $b8_{ ext{S}}$ and variance $c3_{ ext{S}}$ across CARLA simulations and limited real-world tests. The study reveals that factors such as high fog density, heavy precipitation, low sun angles, occlusions, and adversarial modifications to signs cause large predictive-sensitivity and sometimes complete non-detection near $sd$, exposing critical gaps in the current ODDs and perception-planning integration. These results underscore the need for robust perception evaluation, temporal-consistency guarantees, and extended ODD definitions to enhance safety and reliability in real-world AV deployments.

Abstract

The viability of automated driving is heavily dependent on the performance of perception systems to provide real-time accurate and reliable information for robust decision-making and maneuvers. These systems must perform reliably not only under ideal conditions, but also when challenged by natural and adversarial driving factors. Both of these types of interference can lead to perception errors and delays in detection and classification. Hence, it is essential to assess the robustness of the perception systems of automated vehicles (AVs) and explore strategies for making perception more reliable. We approach this problem by evaluating perception performance using predictive sensitivity quantification based on an ensemble of models, capturing model disagreement and inference variability across multiple models, under adverse driving scenarios in both simulated environments and real-world conditions. A notional architecture for assessing perception performance is proposed. A perception assessment criterion is developed based on an AV's stopping distance at a stop sign on varying road surfaces, such as dry and wet asphalt, and vehicle speed. Five state-of-the-art computer vision models are used, including YOLO (v8-v9), DEtection TRansformer (DETR50, DETR101), Real-Time DEtection TRansformer (RT-DETR)in our experiments. Diminished lighting conditions, e.g., resulting from the presence of fog and low sun altitude, have the greatest impact on the performance of the perception models. Additionally, adversarial road conditions such as occlusions of roadway objects increase perception sensitivity and model performance drops when faced with a combination of adversarial road conditions and inclement weather conditions. Also, it is demonstrated that the greater the distance to a roadway object, the greater the impact on perception performance, hence diminished perception robustness.

Figures (17)

• Figure 1: Simple Stopping Scenario
• Figure 2: Experiment Architecture: Data from input sources (left) -- real-time, CARLA, or offline rosbags -- communicated to scalable and extensible AI (right) through ROS framework (center). The AI models perform object classification and detection, deep ensemble processing and additional post processing (right) and communicate back the results.
• Figure 3: Base Experiment: View from the ego vehicle (ego) approaching the stop sign $\mathcal{S}$ in CARLA
• Figure 4: Representative CARLA simulation scenes demonstrating the four distinct occlusion levels used in Experiment Set 1, as viewed from the ego. Each image captures the ego's perspective approaching the stop sign ($\mathcal{S}$) at the same distance and under identical weather conditions, showcasing varying degrees of obstruction by adversarial objects.
• Figure 5: Mean Prediction Probability ($\mu^{Y}_\mathcal{S}$) of YOLOv8 for Stop Sign ($\mathcal{S}$) Detection. Each subgraph displays $\mu^{Y}_\mathcal{S}$ (y-axis) as a function of distance from $\mathcal{S}$ (x-axis) under varying fog densities (0%, 33%, 66%) for a specific occlusion level.