LightDTTrack: light dynamic temporal tracking

In practice, we have found that many methods perform poorly in video sequences where targets are partially occluded or move rapidly, especially under resource-constrained conditions. To address the challenges posed by traditional deep learning-based methods and Transformer-based methods, which hinder effective real-time tracking in specialized scenarios, we propose a tracking algorithm called LightDTTrack. We have simplified the Transformer architecture to significantly reduce the number of parameters and floating-point operations (FLOPs). To maintain effective target position tracking in complex scenes, we introduce a Dynamic Temporal Information Convolution module (DTConv), which enhances the tracker's ability to understand changes in the target across different time frames by incorporating temporal information into the tracking process. Additionally, we have optimized this approach by proposing a fast object screening mechanism. This mechanism screens whether the target is moving rapidly and selectively sends sequence frames to the DTConv module, effectively distinguishing between regular and special scenarios. This helps to further reduce the number of parameters and FLOPs, enabling real-time tracking under limited resources. Finally, we integrate the above modules into a single-object tracking network named LightDTTrack. We conducted experiments on two datasets (OTB100 and UAV123) to validate the effectiveness of our method.