YOLO-STA: A Lightweight Detection Algorithm for UAV Traffic Targets

Abstract This paper proposes YOLO-STA, a lightweight UAV traffic target detection model based on YOLOv12n, aiming to solve practical problems in complex UAV aerial photography scenarios, such as dense targets, large scale variations, strong background interference, and limited computational power of airborne platforms. The model optimizes three aspects—the backbone network, the detection head, and feature enhancement—to adapt to the needs of intelligent traffic monitoring. First, an improved StarNet module is introduced, which utilizes star operation and feature dimension-elevation mechanisms to enhance the representation capability of multi-scale traffic target features, thereby improving the detection accuracy of long-distance and small targets. Second, the Task-Aligned Dynamic Detection Head (TADDH) integrates cross-scale shared convolution, task decomposition and regression-branch DCNv2, with its parameter and inference-cost trade-offs evaluated separately. Finally, the A2C2f_DYT module is adopted, which combines a multi-head regional attention mechanism with a dynamic Tanh activation function to enhance the model's robustness in complex backgrounds and occluded environments through adaptive feature adjustment. The abstract and conclusion report the refreshed comparisons in Tables 4 and 5; Tables 2 and 3 report the earlier ablation batch. The refreshed results indicate that YOLO-STA achieves superior detection performance compared to the baseline model on both VisDrone and VEDAI datasets. On the VisDrone dataset, the precision, recall and [email protected] increase by 0.7, 1.2 and 1.5 percentage points, respectively. On the VEDAI dataset, the [email protected] increases from 63.2% to 66.5%, while the computational complexity decreases from 6.6 GFLOPs to 4.5 GFLOPs, representing a 31.8% reduction. The experimental results demonstrate that YOLO-STA effectively reduces computational complexity while improving detection accuracy, showing a favorable balance between accuracy and complexity. This model possesses significant potential for lightweight deployment in UAV traffic target detection tasks and provides a valuable reference for intelligent traffic monitoring in resource-constrained scenarios.