DilateTracker: Feature Fusion Dilated Transformer for Multi-object Tracking

In recent years, tracking-by-detection (TBD) has emerged as the predominant approach for Multi-object Tracking (MOT). The majority of MOT techniques are either CNN-based or Transformer-based. While CNN is swift, it lacks the ability to model long-range dependencies, whereas the transformer exhibits the opposite characteristics. Several studies have aimed at reducing the computational complexity of the Transformer by modifying the global attention mechanism. For instance, DilateFormer reduces computation costs by employing dilated convolution, yet it lacks interaction between stages. Inspired by this, we have enhanced it by introducing the IDAUP mechanism to facilitate feature interaction between stages. Subsequently, we integrated the lightweight improved DilateFormer into FairMOT as a feature enhancement module, thus enhancing its performance without significantly increasing computational costs. This approach, termed DilateTracker, offers an effective fusion of CNN and Transformer. DilateTracker demonstrates impressive performance on MOT datasets when compared to other advanced methods and is capable of achieving real-time tracking.