Optimization and implementation of lightweight neural networks for real-time object tracking on unmanned aerial vehicles (UAVs)

To address the computational constraints of UAV platforms and challenges like scale variation and occlusion in aerial tracking, this paper proposes LightUAV-Trk—a lightweight and accurate real-time object tracking method. Based on MobileNetV2, it incorporates three key enhancements: an efficient channel-spatial co-attention module to improve feature discrimination for small or occluded objects; a “fuse-before-match” multi-scale fusion strategy that reduces computation by 30% while handling scale changes; and a cosine-similarity-based dynamic template update mechanism to suppress template drift and enhance long-term robustness. With TensorRT FP16 quantization and ROS multi-threading, LightUAVTrk achieves 37 FPS on an NVIDIA AGX Xavier, using only 2.8M parameters and 3.6G FLOPs. On UAV123 and OTB100, it attains 83.9% precision and 65.8% AUC, outperforming leading lightweight trackers such as SiamRPN and ConcatTrk, especially under occlusion and scale variation, offering a practical solution for UAV embedded vision tasks.