Abstract: The capability to detect small objects in complex road environments is crucial for the safety and robustness of autonomous driving systems. Especially under adverse weather conditions such as low illumination and fog, the images captured by visual sensors suffer from reduced contrast, blurred details, and unclear object boundaries due to insufficient lighting, droplet occlusion, and atmospheric scattering effects, which easily lead to missed and false detections. To address this challenge, this paper proposes a small object detection method guided by visual features for complex environments, from three aspects: training data construction, network architecture design, and sample allocation strategy. Firstly, to overcome the lack of nighttime foggy training data, a depth-aware atmospheric scattering physical model is designed based on the clear-weather KITTI dataset. This model realistically simulates the scattering and attenuation of light in fog by considering scene depth, fog density, and illumination intensity, and introduces a low-illumination rendering strategy to generate diverse and realistic nighttime foggy images. This expanded dataset significantly improves the model’s generalization ability under extreme weather conditions. Secondly, on the detection network side, a Multi-Layer Channel Fusion Module (MLCFM) is proposed based on the YOLOv11 framework. By splitting, reorganizing, and adaptively weighting feature channels at different levels, the module preserves low-level texture details while enhancing high-level semantic discrimination, effectively extracting critical features for small objects.Then, a semantics-importance-driven dynamic multi-scale fusion structure is designed. It dynamically adjusts fusion weights based on the semantic contribution of features at different scales to various object categories, strengthening perception of small-sized targets such as pedestrians and cyclists, while maintaining global contextual information for larger targets like vehicles.