AL-YOLO: Accurate and Lightweight Vehicle and Pedestrian Detector in Foggy Weather

Object detection remains a crucial challenge in computer vision, particularly under adverse weather conditions such as fog, where reduced visibility significantly impacts detection accuracy. Many existing methods, including image restoration and domain adaptation techniques, attempt to address these issues but often result in increased network parameters or negative transfer effects, limiting their practicality on resource-constrained devices. In this study we introduce AL-YOLO designed specifically for effective vehicle and pedestrian detection in foggy weather. By adapting YOLOv5s architecture, we integrate the MobileViTv2 with an Inverted Residual Block as the backbone, optimizing computational efficiency and enabling robust feature extraction. Additionally, we incorporate a C3RFE module to expand the receptive field, which significantly enhances detection performance in low-visibility conditions. Our experimental results on the Foggy-Cityscape dataset demonstrate that AL-YOLO achieved a 33% reduction in model parameters and a 2% mean Average Precision (mAP) improvement over YOLOv5s, emphasizing its efficiency and suitability for real-time fog-resistant object detection on resource-limited devices. Code and pretrained models are available at: https://github.com/BehdadSDP/AL-YOLO.

I. Introduction

Object detection is a fundamental task in computer vision with applications in diverse fields such as autonomous driving, surveillance, and robotics. However, most current object detection models face challenges when operating in adverse weather conditions like fog and haze, where visibility deteriorates, significantly reducing detection accuracy [1]. To address this challenge, researchers have proposed various solutions, including image restoration techniques, which enhance visibility by applying dehazing as a preprocessing step [2]. Utilizing image restoration as an initial step before detection can enhance visibility. However, executing separate dehazing and detection models consecutively leads to increased computing costs and reduced inference speed. This presents challenges for systems with limited resources that require real-time performance. Moreover, optimizing restoration and detection separately does not result in a collaborative improvement of both tasks in an integrated manner [3]. Furthermore, dehazing can accidentally eliminate important features that are crucial for subsequent detection processes [4], [5].