Progressive Hierarchical Feature Periodic Transformer for Robust UAV Tracking

Abstract

Robust visual tracking for unmanned aerial vehicles (UAVs) remains challenging in dynamic aerial environments because existing methods either fail to capture long-range dependencies in CNN-based models or lose critical local details in Transformer-based approaches. To overcome these limitations, we propose the Progressive Hierarchical Feature Periodic Transformer (PHFP), a two-stage representation learning framework for UAV tracking. In the coarse representation learning stage, our Progressive Hierarchical Feature Refinement (PHFR) module employs a multi-branch atrous Feature Enhancement Module (FEM) to enrich fine-grained local contexts, along with Semantic-Aware Module (SAM) and Parallel Asymmetric Convolution Block (PACB) that leverage shallow features to suppress background noise and refine deep representations. In the fine representation learning stage, a periodic transformer, enhanced by a dedicated Fourier Analysis Network (FAN), hierarchically fuses multi-level feature maps to explicitly extract latent periodic structures and generate robust embeddings. Extensive evaluations on the DTB70, UAV20L, UAVTrack112, and UAVTrack112_L benchmarks demonstrate that PHFP outperforms state-of-the-art trackers by significant margins while operating at 32.6 frames per second on an edge-equipped aerial platform, confirming its effectiveness and real-time capability.