Abstract: To address issues such as false positives, missed detections, and poor detection performance in small object detection from drone imagery, an improved RTDETR model is proposed. The backbone network is enhanced with a BasicAK MECS module, which integrates a lightweight adaptable kernel convolution (AKConv) and a median-enhanced spatial channel attention mechanism, thereby improving feature extraction. The Neck component incorporates Slim-neck to both lighten the model and increase detection accuracy and efficiency. In the upsampling stage, a Dynamic Scale Sequence Feature Fusion (DySSFF) module is utilized to enhance the extraction of small object features. The loss function employed is Inner-SIoU, which incorporates shape and scale information of bounding boxes to improve bounding box regression performance. Experimental results demonstrate that the proposed model achieves a 2.2% increase in mAP and a 0.9% increase in mAP_s on the Visdrone2019 dataset, significantly enhancing the model's object detection capability.