Based on low-cost and high-information-content ship visual image data, this paper utilizes deep learning and machine vision technologies to extract navigation information of ships from images. This research is crucial for enhancing ship transportation efficiency, ensuring marine traffic safety, promoting environmental protection, and optimizing ship supervision. To address these issues, a machine learning-based framework for ship speed extraction in foggy environments is designed. A multi-scale AOD-Net network structure is constructed to improve network performance, strengthening the network's ability to process detailed features and enhance the recovery of local details of small targets in defogged images. Using YOLOv5-Deep SORT algorithm, moving ships in defogged images are detected and tracked from maritime video images, and ship trajectory information between video frames is output to calculate pixel displacement of ships. Finally, the actual ship speed is estimated through the 2D-3D spatial mapping relationship. Experiments show that the framework achieves an average accuracy of approximately 95% in extracting ship speeds in both simulated and real-world scenarios. The mean squared error (MSE) of speed values extracted from defogged images is approximately 0.3 Kn lower than that from images before defogging, ensuring high accuracy in ship speed extraction.