The multi-object tracking of inland waterway vessels is an important and challenging task for intelligent monitoring of waterway traffic. Compared to tracking pedestrians and vehicles, inland vessels exhibit significant scale variations, longer tracking duration, and are more susceptible to occlusion in video surveillance. To address these challenges, we propose a dual-space approach for multi-object tracking of inland vessels, based on both image and geographic spaces. Firstly, we introduce a boundary attention mechanism to tackle the problem of texture and color information loss during the scale variation in vessel images. Next, leveraging the principles of homography transformation, we convert the vessel position from image coordinates into geographic coordinates, overcoming the difficulty of obtaining accurate motion information caused by scale variations in vessel images. Finally, we propose a dual-space distance metric to measure the similarity of vessel objects between adjacent video frames, enabling real-time matching and tracking of multiple vessels. Through experimental analysis in various surveillance scenarios, such as navigation in curved waterways and entering/exiting anchorage areas, our method is validated. Compared to existing methods like Deep-Sort and Strong-Sort, our approach effectively mitigates the impact of vessel image scale variations and occlusion. It demonstrates significant improvements in accuracy, recall rate, and identity switch frequency.