Abstract: Object detection in UAV aerial imagery often faces challenges such as large model size, high parameter count, and limited detection accuracy, which restrict its deployment on embedded and edge devices with constrained computing power and low energy budgets. To address these issues, GSC-YOLO, a lightweight object detection model based on an improved YOLOv5 architecture is proposed. Specifically, ghost convolution is introduced into the backbone network to significantly reduce the number of parameters and overall model size. A simple attention mechanism is also integrated to enhance the network’s ability to extract critical features by leveraging global contextual information. Furthermore, the traditional non-maximum suppression （NMS） is replaced by the more efficient Cluster-NMS, and additional improvements such as score penalty and center-distance mechanisms are applied to further refine detection results. Experiments on the VisDrone2019 dataset demonstrate that GSC-YOLO achieves a 3.1% improvement in mean average precision （mAP）, with an 82.1% reduction in parameters and a 77.4% decrease in model size compared to the baseline YOLOv5s. These results show that GSC-YOLO offers a better trade-off between accuracy and efficiency than other mainstream YOLO variants and lightweight models. Its successful deployment on the KV260 platform further validates its practicality for real-time UAV applications on embedded and edge devices.