Parallel Perceptual Attention and Multifrequency Matching Network for UAV Tracking

Abstract

In recent years, trackers based on deep learning have demonstrated immense potential, attributed to their robust modeling capabilities. Nevertheless, due to the limited computational power of uncrewed aerial vehicle (UAV) platforms and the complexity of scenarios encountered during tracking, the existing trackers struggle to effectively balance algorithmic accuracy and operational speed. This defect seriously affects the practical significance of tracking algorithm based on deep learning. To address this challenge, we have devised PMTrack, an efficient tracking model grounded in Siamese neural networks. The key innovations of PMTrack encompass: 1) the adoption of parallel perceptual attention (PPA) to enhance feature saliency and 2) the design of a multifrequency matching (MFM) network that facilitates feature matching through multidimensional feature information while mitigating redundant computations. PMTrack is both efficient and effective: its effectiveness is validated through comprehensive evaluations on multiple public benchmarks. We have deployed PMTrack on various drone platforms, specifically running at a speed of 46 frames per second (FPS) on the typical embedded aerial tracking platform Nvidia Xavier. This confirms the feasibility and practicality of the algorithm presented in this letter in real world.