PaniCar: Securing the Perception of Advanced Driving Assistance Systems Against Emergency Vehicle Lighting

Elad Feldman 1 , Jacob Shams 1 , Dudi Biton 1 , Alfred Chen 2 , Shaoyuan Xie 2 Satoru Koda 3 , Yisroel Mirsky 1 , Asaf Shabtai 1 , Yuval Elovici 1 , Ben Nassi 1 · 2025

Read PaniCar: Securing the Perception of Advanced Driving Assistance Systems Against Emergency Vehicle Lighting by Elad Feldman 1 , Jacob Shams 1 , Dudi Biton 1 , Alfred Chen 2 , Shaoyuan Xie 2 Satoru Koda 3 , Yisroel Mirsky 1 , Asaf Shabtai 1 , Yuval Elovici 1 , Ben Nassi 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-05183.txt.

Introduction

Elad Feldman and nine co-authors expose a hazard in 2025 for driver-assistance vision: flashing emergency lights. Sixteen crashes of Teslas on autopilot into parked emergency vehicles frame the study. Bright strobes splash lens flare across camera images. The team shows the flare makes object detectors' confidence flicker below their trigger lines, so real objects near ambulances can vanish from the car's view.

They name the effect PaniCar and test it across seven commercial systems, four detector families, and fourteen light patterns. Standard flare-removal tools prove too slow and too weak. Their defense framework, Caracetamol, hardens detectors against the flicker.

Catalog record

CREATORElad Feldman 1 , Jacob Shams 1 , Dudi Biton 1 , Alfred Chen 2 , Shaoyuan Xie 2 Satoru Koda 3 , Yisroel Mirsky 1 , Asaf Shabtai 1 , Yuval Elovici 1 , Ben Nassi 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-05183

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

PaniCar: Securing the Perception of Advanced Driving Assistance Systems Against Emergency Vehicle Lighting Elad Feldman 1 , Jacob Shams 1 , Dudi Biton 1 , Alfred Chen 2 , Shaoyuan Xie 2 Satoru Koda 3 , Yisroel Mirsky 1 , Asaf Shabtai 1 , Yuval Elovici 1 , Ben Nassi 1 Abstract The safety of autonomous cars has come under scrutiny in recent years, especially after 16 documented incidents involving Teslas (with autopilot engaged) crashing into parked emergency vehicles (police cars, ambulances, and firetrucks). While previous studies have revealed that strong light sources often introduce flare artifacts in the captured image, which degrade the image quality, the impact of flare on object detection performance remains unclear. In this research, we unveil PaniCar , a digital phenomenon that causes an object detector’s confidence score to fluctuate below detection thresholds when exposed to activated emergency vehicle lighting. This vulnerability poses a significant safety risk, and can cause autonomous vehicles to fail to detect objects near emergency vehicles. In addition, this vulnerability could be exploited by adversaries to compromise the security of advanced driving assistance systems (ADASs). We assess seven commercial ADASs (Telsa Model 3, "manufacturer C", HP, Pelsee, AZDOME, Imagebon, Rexing), four object detectors (YOLO, SSD, RetinaNet, Faster R-CNN), and 14 patterns of emergency vehicle lightning to understand the influence of various technical and environmental factors. We also evaluate four SOTA flare removal methods and show that their performance and latency are insufficient for real-time driving constraints. To mitigate this risk, we propose Caracetamol , a robust framework designed to enhance the resilience of object detectors against the effects of activated emergency vehicle lighting. Our evaluation shows that on YOLOv3 and Faster R-CNN, Caracetamol improves the models’ average confidence of car detection by 0.20, the lower confidence bound by 0.33, and reduced the fluctuation range by 0.33. In addition, Caracetamol is capable of processing frames at a rate of between 30-50 FPS, enabling real-time ADAS car detection. I Introduction The safety of autonomous cars has been questioned in recent years following 16 documented incidents [ 1 ] in which Tesla cars with engaged autopilot collided with parked emergency vehicles, such as fire engines [ 2 ] , police cars [ 3 ] , and ambulances [ 4 ] . These accidents resulted in significant injuries and fatalities [ 5 ] , causing the National Highway Transportation Safety Administration (NHTSA) to launch an in-depth investigation of the incidents to determine the factors underlying these accidents [ 6 ] . While driver distraction has been identified as a primary cause [ 5 ] , the technical aspects that led to these accidents remain a mystery [ 5 ] . Fig. 1: Top: A vehicle detected by SSD with a confidence of 0.98 with emergency vehicle lighting off (left) and 0.03 with emergency vehicle lighting on (right) in images obtained by a 2023 Tesla model 3 frontal camera. Bottom: The associated normalized tonal distribution of the images. Despite public interest and the importance of the investigation, neither the automotive industry nor NHTSA has disclosed their findings, leaving critical questions unanswered regarding the technological factors behind these incidents. Prior studies [ 7 , 8 , 9 ] have demonstrated that strong light sources—such as those from street lamps—can introduce flare artifacts, degrading image quality by lowering metrics like peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). These studies establish that flare degrades the visual quality of single images captured by standard video cameras in the presence of static light sources. However, the incidents in question involved advanced driving assistance systems (ADASs) cameras (rather than standard video cameras), failures in perception and object detection over time (rather than quality degradation in a single frame), and emergency vehicle lighting, which emits variable-intensity light in dynamic patterns (unlike constant-intensity street lamps). The gap between the conditions studied in prior work [ 7 , 8 , 9 ] and the real-world conditions relevant to the documented failures leaves a critical question unaddressed: What is the impact of flare from emergency vehicle lighting on object detection performance over time in videos containing fire trucks, ambulances, and police cars? This question must be answered, especially in light of the increased adoption of semi-autonomous cars and ADASs, for multiple reasons: (1) Such accidents can result in injuries or the death of drivers, passengers, and pedestrians [ 5 ] . (2) There is an industry shift in the architectural design of new semi-autonomous cars that is increasingly leaning toward vision-based systems [ 10 ] without RADARs and ultrasonic sensors. (3) Until these questions are answered, the ability to develop methods to secure the perception of semi-autonomous cars and other perception-based systems (e.g., drones, humanoids) will be curtailed, exposing the perception of autonomous vehicles to unintentional safety faults caused by emergency vehicle lightning or intentional security attacks performed by adversaries. This study aims to explore the role of object detectors in the failure of vision-based perception when encountering a vehicle with activated emergency vehicle lighting. This is not only important from a safety perspective, but also from a security perspective, since attackers can leverage this failure to target semi-autonomous cars and deliberately cause damage and endanger the lives of road users. First, we analyze video footage obtained from two commercial Level 2 ADASs (Tesla Model 3 and "manufacturer C") and five commercial Level 1 ADASs (HP [ 11 ] , Pelsee [ 12 ] , AZDOME [ 13 ] , Imagebon [ 14 ] , Rexing [ 15 ] ) and show that four object detectors are vulnerable to a phenomenon we have termed PaniCar , a digital fluctuation in the confidence score of object detectors of a detected object caused by the activation of emergency vehicle lighting. This has dangerous implications, as the fluctuation in the confidence score may result in a value below the detection threshold, causing an object to go unnoticed ( i.e., undetected) by semi-autonomous cars and ADASs. In our investigation of the PaniCar phenomenon, we discover that the confidence score of an object detector is affected by the intensity of the light emitted from activated emergency vehicle lighting. The added flare from the activated emergency vehicle lighting changes the tonal distribution of the car (in the frame captured by the video camera) over time (between different frames) according to the pattern and frequency of the activated emergency vehicle lighting. Consequently, lower confidence detection scores are produced by the object detector in different frames, affecting the object detector’s behavior and performance and causing detection loss of objects that appear in the frame (see Fig. 1 ). We also investigate the factors affecting the PaniCar phenomenon and find that it appears in a variety of object detectors and commercial ADASs, creates unique detection behavior depending on the pattern of the emergency vehicle lighting, and mainly compromises object detectors’ performance in dark conditions. To address the PaniCar phenomenon, we evaluate the effect of state-of-the-art (SOTA) flare removal methods [ 7 , 8 , 9 ] on the performance of object detectors and find a scientific gap: while SOTA flare removal methods improve the quality of the picture (the PSNR and SSIM), they do not improve the performance of the object detectors significantly

Read the complete text