Aiming at the problems such as low accuracy and poor robustness of model detection due to fuzzy image in the process of target detection in foggy scenarios, the YOLOv5 algorithm is optimized and improved in combination with data enhancement, and a target detection method based on dimensional interaction and cross-layer scale cascading is proposed. Firstly, triplet attention is embedded in the feature extraction structure to capture the dependencies between different dimensions, enhance the fusion and interaction of information between spaces and channels, and improve the ability to focus on important features. Secondly, a Multi-scale Receptive-field Enhancement Module (MREM) is proposed. Multiple repeated pooled sampling and residual connection were used to enlarge the target receptive field to obtain multi-scale features and enhance the ability of the model to extract details. Thirdly, the structure of Cross-Layer Cascading Path Aggregation Network(CLC-PAN) is proposed. Cross-layer connection is adopted to promote the fusion of feature information of different scales, improve the interaction between shallow detail information and deep semantic information, and capture richer semantic features by deepening the sampling layers of the feature pyramid, so that the laying intervals of various anchor frames are more reasonable and the model detection ability is improved. Finally, the SIoU loss function is used as the target bounding box regression loss function to improve the target box positioning accuracy and sample training speed. The experimental results show that the model size of the improved detection method is 15.8 MB, and the mAP reaches 71.3%, which is 7% higher than that of YOLOv5s, which can meet the needs of fast and accurate real-time target detection in foggy scenes.