Edge-Assisted Object Perception for Autonomous Vehicles Under Challenging Exposure and Blur Conditions

With the rapid advancements in machine learning (ML) and computing power, we have witnessed extensive deployment and testing of autonomous vehicles globally. By deploying ML models tailored for different scenarios on edge devices, including connected and autonomous vehicles, they have become increasingly intelligent in handling tasks such as object detection, lane keeping, and pothole detection. However, significant safety concerns persist, particularly concerning object detection in corner cases. We have identified two critical scenarios that urgently require effective handling in camera-based autonomous driving systems: challenging exposure and blur conditions. In this paper, we design and implement a new YOLO-EXB model based on the YOLO structure to enhance the safety of autonomous driving. First, we collect abnormal images for these two scenarios to compensate for the lack of corner case data. Next, we propose a novel backbone architecture integrating Transformer components into the original design, enhancing its feature extraction capabilities and modeling power. Finally, we evaluate our YOLO-EXB model on testing images under those two challenging conditions. Aiming to mitigate the impact of strong light and highly blurred images on autonomous driving perception, the proposed YOLO-EXB model has demonstrated improved detection capabilities and enhanced robustness under challenging exposure and blur conditions compared to both the baseline YOLOv8 and state-of-the-art YOLO11 models. In the conclusion of this paper, we discuss our limitations and highlight our contributions to improving the safety and effectiveness of autonomous driving in corner-case scenarios.