End-to-end wavelet block feature purification network for efficient and effective UAV object tracking

Recently, unmanned aerial vehicle (UAV) object tracking tasks have significantly improved with the emergence of deep learning. However, owing to the object feature pollution caused by motion blur, illumination variation, and occlusion, most of the existing trackers often fail to precisely localize the target in the complex real-world circumstances. To overcome this challenge, we present a novel wavelet block feature purification network (WFPN) for efficient and effective UAV tracking. WFPN is mainly composed of downsampling network through wavelet transforms and upsampling network through inverse wavelet transforms. To be specific, the downsampling network performs discrete wavelet transform (DWT) to reduce interference information and preserve original feature details, while the upsampling network applies inverse DWT (IDWT) to reconstruct decontaminated feature information. Additionally, a novel sequential encoder is introduced to achieve a better purification effect. Finally, a pooling distance loss is devised to improve the purification effect of DWT downsampling network. Extensive experiments show that our WFPN achieves promising tracking performance on three well-known UAV benchmarks, especially on sequences with feature pollution. Moreover, our method runs at 33.2 frames per second on the edge platform of Nvidia Jetson AGX Orin, which is suitable for UAVs with limited onboard payload and computing capability.