<p>Unmanned Aerial Vehicle (UAV) edge detection platform has the disadvantages of limited computing resources and storage capacity. To better complete autonomous tasks in a complex and changeable background, efficient and lightweight detection algorithms are still needed to meet the dual requirements of real-time performance and accuracy. The existing lightweight detectors are often accompanied by significant increases in the number of parameters and computational overhead when the accuracy improves, making it difficult to deploy efficiently on resource-constrained devices. To reduce the computing load and improve the detection performance of unmanned aerial vehicles (UAV) in edge scenarios, we propose a lightweight object detection algorithm CFA-YOLOv10 by combining the C3RFEM feature extraction module, the FreqFusion feature fusion module, and the AIFI self-attention module. Firstly, we introduced the C3RFEM module to replace the C2f module in the backbone network. We utilized multi-branch expanded convolution and residual connections to expand the effective receptive field, while reducing the number of parameters of the feature extraction structure by 7.2%. Secondly, the FreqFusion module is introduced into the neck network to enhance semantic consistency and sharpen boundaries through adaptive low-/high-pass filters and offset generators, effectively alleviating the problems of intra-class differences and boundary dislocations. Next, the PSA structure is replaced with the AIFI module, and its serialized self-attention mechanism is utilized to enhance the expression of advanced semantic information and reduce the loss of key features. Finally, the Shape-NWD loss function replaces the traditional loss function, integrating shape weighting coefficients with normalized distance metrics to enhance the regression accuracy of bounding boxes for small targets. A large number of experiments on the VisDroneDET-2019 and UAVDT datasets show that CFA-YOLOv10 only requires 7.88M parameters and 17.9 GFLOPS of computational cost, achieving real-time performance of 39.1% mAP50 (an increase of 0.3% compared to the original YOLOv10-S) and 180 FPS. It is significantly superior to other mainstream lightweight models, providing a high-precision and low-latency detection solution for the edge platform of unmanned aerial vehicles. More importantly, this algorithm is specifically designed for edge high-performance computing scenarios. Through operator-level optimization and computational graph reconstruction, on the NVIDIA Jetson AGX Orin edge computing module, the CFA-YOLOv10 achieves 180 FPS high-throughput inference at 22.8W power consumption, with a GPU utilization rate of 92%. The energy efficiency ratio reached 7.89 FPS/W, a significant improvement over the benchmark model. This algorithm provides a high-precision, low-latency, and high-efficiency detection solution for resource-constrained unmanned aerial vehicle (UAV) edge platforms.</p>