Lightweight Object Detection for Traffic Monitoring in Adverse Weather Using a Case Study in Jeju

This study proposes four object detection models that integrate the latest backbones with attention mechanisms and comprehensively compares and analyzes their performance in adverse weather conditions. Adverse environments such as fog, rain, and low light degrade video quality, necessitating lightweight technology for effective object detection using only existing CCTV infrastructure, especially in budget-limited regions. We propose and compare the performance of four lightweight models based on YOLOv8 and YOLOv11. The first two models combine Swin Transformer and the EMA (Efficient Multi-scale Attention) module to leverage global contextual information, while the latter two apply the EfficientNet backbone to substantially reduce complexity and maintain high accuracy. Experiments using the RTTS and Fog or Detection datasets, along with Jeju Traffic CCTV footage, showed that YOLOv11-EfficientNet-EMA achieved the most superior performance, reaching 57.90% based on mAP50-95. However, the models face limitations like performance degradation in extremely heavy rain or completely unlit nighttime conditions, requiring minimal edge computing equipment for real-time processing. The proposed lightweight models can be applied to existing traffic CCTV infrastructure through software updates alone, serving as a low-cost, high-efficiency solution for traffic monitoring in fog-prone areas like Jeju, real-time accident detection, and future autonomous-driving-based traffic safety systems, and are expected to contribute to infrastructure development in rural areas or developing countries where expensive sensor deployment is difficult.