Enhanced Object Detection in Adverse Weather Conditions Using YOLOv8

Object detection in adverse weather conditions presents significant challenges due to dynamic objects and continuous environmental changes. Weather phenomena such as snow, rain, fog, low light, sandstorms, and intense sunlight often degrade the quality of perception, complicating accurate detection. In this work, it aims to improve camera-based object detection accuracy under harsh weather conditions. We propose enhancing YOLOv8-based object detection through transfer learning, utilizing merged data from multiple harsh weather datasets. Two prominent open-source datasets, ACDC and DAWN, along with their merged versions, were employed to detect primary road objects in challenging conditions. A set of training weights was generated from training on individual datasets, their merged counterparts, and several subsets categorized by specific characteristics. A comparative analysis of these training weights was conducted by evaluating detection performance across the datasets and their subsets. The results demonstrated that training with custom datasets significantly improved detection performance compared to the YOLOv8 base weights with 85.6% precision, 83.1% recall value, and 84.5% mean average precision. Also provides the better outcomes thsn the benchmark models. Additionally, incorporating more images through a feature-related data merging technique consistently enhanced object detection accuracy, highlighting the potential for further optimization in adverse weather conditions.