<p>Occlusion, sudden illumination changes, and rapid motion in complex scenes severely degrade the robustness of existing object tracking methods. To address this issue, this paper proposes a novel object tracking algorithm that integrates a deformable attention mechanism. The method first embeds a deformable attention module into the ResNet-18 feature extraction network to enable adaptive enhancement of target key features. Second, the method adopts an improved Bidirectional Feature Pyramid Network as the feature fusion module to enhance the representational capability of multi-scale features. Finally, the method incorporates a dynamic Kalman filtering prediction module to improve the algorithm’s adaptability to changes in the target’s motion state and its continuous tracking capability. Experimental results show that the improved feature extraction network achieves an average overlap rate and success rate of 61.5% and 68.4%, respectively, on the GOT-10k dataset, with a computational load of only 1.96 GFLOPs and an increase of only 0.23&#xa0;M in parameters. On the MOT20 dataset, the proposed object tracking network achieves a Multiple Object Tracking Accuracy of 77.5%, an Identity F1 Score of 77.0%, with 54.6% Majority of Tracked Trajectories and 12.5% Majority of Lost Trajectories. Its tracking performance surpasses that of the compared object tracking algorithms. These results confirm the efficacy of the Deformable Attention Mechanism and present a robust solution for complex dynamic tracking scenarios.</p>