Feature-enhanced and lightweight object detection method for UAV perspective
Abstract
Unmanned aerial vehicle (UAV) target detection faces significant challenges such as extreme scale variations, complex background interference, and limited computational resources on mobile edge devices. To address the high miss rate of small objects and high computational redundancy, this paper proposes FSD-YOLO, a lightweight and feature-enhanced target detection algorithm based on YOLO11. First, a lightweight architecture combining a FasterNet backbone and a BiFPN neck is constructed to reduce computational costs while enhancing the bidirectional fusion of cross-scale features. Second, a Local Super-Resolution Convolution (LSRConv) module is proposed to reconstruct the fine-grained texture details of tiny objects using a space-channel-space transformation paradigm, effectively mitigating critical feature loss during continuous downsampling. Finally, a lightweight DEfficientHead-P2 decoupled head is designed. By integrating a P2 micro-scale detection layer and optimizing with depthwise separable convolutions, the network expands its receptive field for tiny targets while significantly compressing model parameters. Experimental results on the VisDrone2019 dataset demonstrate that the proposed method achieves an mAP50 of 37.6% with a computational load of only 5.8 GFLOPs, outperforming the baseline YOLO11n model. Furthermore, tests on the TinyPerson dataset verify its robust generalization in search-and-rescue and maritime surveillance scenarios. Real-world edge deployment on the NVIDIA Jetson Xavier NX achieves an inference speed of 98 FPS with a low power consumption of 11.6W. The proposed algorithm achieves an excellent balance between detection accuracy and inference efficiency, making it highly suitable for real-time target perception on resource-constrained embedded edge systems.