To address the demand for high-precision and lightweight object detection in low-altitude traffic scenarios with unmanned aerial vehicles (UAVs), this study proposes an improved detection model, namely the Cross-scale Alignment and Positional Encoding-enhanced RT-DETR (CAPE-RT-DETR), based on the Real-Time Detection Transformer (RT-DETR). First, a feature enhancement module integrating dynamic convolution kernel generation and gated feature selection is designed to strengthen the extraction and filtering of key features. Second, a position-aware interaction module is constructed by combining learnable positional encoding with the multi-head attention mechanism, which effectively enhances the model’s perception of spatial structures and localization accuracy. Finally, a pyramid scene parsing structure is introduced to integrate sparse global contextual information, while a dual convolution and grid sampling mechanism is employed to explicitly compensate for cross-scale feature alignment deviations, thereby avoiding misalignment caused by upsampling operations and further improving detection performance. Experiments conducted on the ALU and VisDrone2019 datasets, with comparisons against 18 state-of-the-art detection methods, demonstrate that CAPE-RT-DETR outperforms the baseline in terms of parameter efficiency, accuracy, and model size. In addition, ablation studies validate the effectiveness and complementarity of the three proposed modules. This work provides a high-precision and lightweight algorithmic foundation and theoretical support for real-time UAV object detection in complex scenarios.