Research on Improved Small-Object Detection Algorithm for UAVs from Low-Altitude Perspectives

In the context of the rapidly growing low-altitude economyand the urgent need for high-precision, lightweight UAV object detection in agriculture, logistics, and emergency rescue, this paper tackles the challenges of low target pixel occupancy, environmental occlusion, and severe perspective distortion in low-altitude imagery. Based on the RT-DETR algorithm, an enhanced detection model—Cross-scale Alignment and Position Encoding Enhanced RT-DETR (CAPE-RT-DETR)—is proposed. Firstly, to overcome the limitation of traditional static convolution kernels in terms of feature extraction flexibility under complex backgrounds, this paper proposes a feature enhancement moduleintegrating dynamic convolution kernel generation and gated feature selection, termed C2ML. By utilizing a Large Kernel Predictor (LKP) to dynamically generate spatially adaptive convolution kernels, and combining it with a gated feature selection mechanism to eliminate redundant background information, the module significantly enhances the model’s ability to extract and filter critical features. Secondly, to address the geometric distortion and spatially non-uniform deformation characteristic of aerial perspectives, a learnable positional encoding is integrated with the multi-head self-attention mechanism to construct an enhanced position-aware interaction module, termed AIFP. By learning spatial prior information in an end-to-end manner, this module effectively improves the model’s perceptual sensitivity and localization accuracy with respect to the low-altitude-specific spatial structures. Finally, to resolve the pixel misalignment problem caused by simple upsampling in multi-scale feature fusion, a a cross-scale feature calibration (CSFC) module is introduced. This module utilizes a pyramid scene parsing structure to integrate sparse global context and employs a dual-path convolution and grid sampling mechanism to explicitly compensate for cross-scale alignment biases, thereby achieving consistent representation of semantic information. Experimental results on the ALU and VisDrone2019 datasets demonstrate that CAPE-RT-DETR outperforms the baseline algorithm in terms of parameter count, accuracy, and model size. Meanwhile, ablation experiments validate the effectiveness and synergy of the three improved modules. This research provides a high-precision and lightweight methodological foundation and theoretical support for real-time UAV object detection in complex scenarios.