Research on UAV personnel detection in complex environments based on improved YOLOv11

We propose an improved model, YOLO-CED, built upon YOLOv11n to address the persistent challenges of low detection accuracy and high miss rates in unmanned aerial vehicle–based personnel detection under complex environments. The primary contribution lies in the design of a new lightweight backbone, which combines MobileNetV3 with existing YOLOv11 modules: spatial pyramid pooling fast and convolutional block, with parallel spatial attention to significantly reduce parameter count and computational cost while maintaining high detection accuracy. In addition, a profound semantic fusion module is introduced in the neck to perform deep semantic fusion across multiscale features using cross-attention, enhancing the model’s capacity to understand complex spatial information. A useful task dynamic aligned detection head is also developed to decompose and align detection tasks more effectively through dynamic convolution and adaptive feature selection while sharing features to minimize redundancy. Moreover, an inner-complete intersection over union loss function is employed to improve the precision of small-object localization. Experimental evaluations demonstrate that YOLO-CED achieves a strong trade-off between lightweight architecture and detection performance. Specifically, it improves precision (86.75%), recall (79.61%), and mAP50 (87.57%) by 2.76%, 2.14%, and 3.97%, respectively, over the YOLOv11n baseline, and performs competitively against other mainstream detectors under similar computational constraints. This work is essential for real-time aerial surveillance and emergency response applications, offering a practical and efficient solution for robust personnel detection in complex scenarios.