Research on lightweight algorithm of UAV infrared pedestrian detection based on YOLOv11n

Aiming at the poor quality of infrared images collected by UAVs in low-visibility environments such as night, and the contradiction between the lightweight deployment of existing detection algorithms on edge devices and detection accuracy, this paper proposes a balance strategy of “structure simplification + attention compensation” based on YOLOv11n as the baseline model. Firstly, the Slim-neck module is reconstructed, and depthwise separable convolution (DWConv) is used to replace the standard convolution. The original dual-branch CSP structure is simplified to a single-branch EfficientVoVGSCSPC module. At the same time, the complex tensor rearrangement operation is removed, which significantly reduces the parameters, GFLOPs, and size of the model. Then, in order to make up for the decline in feature expression ability caused by lightweight design, an ECA module is embedded in the improved GSConv to achieve local enhancement of channel features. In addition, ECA and CBAM are cascaded at the output of EfficientVoVGSCSPC to form a dual-attention module, and the multi-level features are serially refined in the order of “channel-level lightweight calibration → space-level fine weighting”. The experimental results on the self-built UAV infrared pedestrian dataset show that the final model maintained the detection accuracy (mAP@0.5 of 87.5%) while reducing the parameters (42.6%), GFLOPs (28.6%), and model size (39.2%). Compared to the baseline model, although the inference frame rate has decreased by approximately 36%, it offers significant advantages in terms of model parameters, GFLOPs, and size. In addition, the effectiveness of the ECA → CBAM sequence and the advantages of the “single-branch + attention” configuration are verified by comparing different attention cascade sequences and the interaction between residual and attention. This method offers a potential reference for the lightweight design of UAV infrared pedestrian detection, though further validation on actual edge hardware is needed.