With the increasing demand for the application of intelligent mobile robots in complex environments, dynamic object tracking technology has become an important task for robot autonomous perception. However, traditional tracking algorithms often face problems of reduced accuracy and insufficient robustness when dealing with complex backgrounds, rapid motion, and occlusion interference. Therefore, a deep learning-based dynamic object tracking algorithm is proposed. Experimental results demonstrate that SAFNet exhibits excellent performance in various dynamic environments. Under conditions of low speed (0.2 m/s) and low occlusion (10%), SAFNet achieves a tracking accuracy of up to 83.4% and a success rate of 89.5%. In high-speed (0.8 m/s) and high-occlusion (30%) environments, SAFNet still maintains an accuracy of 74.3% and a success rate of 78.8%. Compared with traditional algorithms, SAFNet significantly reduces the object loss rate. Under conditions of high illumination changes and complex occlusions, the object loss rate is as low as 5.7%, far outperforming SiamFC (27.5%), SiamRPN++ (21.3%), and Ocean (18.9%). SAFNet has significant advantages in dynamic object tracking, effectively improving tracking accuracy and stability in complex environments and under high motion speeds, providing stronger perception capabilities and application potential for mobile robots.