Abstract: Aiming at the technical bottlenecks such as the insufficient detection accuracy of multi-scale targets and the low recognition rate of small targets in the large-view dynamic scenes of tower videos,this paper proposes a ADM-YOLOv11 target detection algorithm based on the collaborative optimization of multiple modules.Firstly,the adaptive feature enhancement(AFE)module is embedded in the Backbone network to deeply reconstruct the C3K2 module.Through the spatial context awareness and feature refinement mechanisms,the feature extraction ability of the network for complex scenes is significantly improved.Secondly,the efficient multi-scale attention (EMA)module is integrated into the C3K2 module of the Neck to enhance the detection robustness of the model for multi-scale targets.Thirdly,the ultra-lightweight dynamic upsampler DySample (dynamic upsample)is introduced into the Neck structure to replace the traditional upsampling layer,optimizing the detail expression and semantic fusion efficiency of multi-scale features.Finally,the EMASlideLoss classification loss function is adopted.By using a dynamic weighting strategy,the problem of gradient shift caused by data imbalance is suppressed,effectively improving the generalization performance of the model.The experimental results show that for the model in this study,the mAP_50-95 is increased from 74.8%of the baseline model to 82.6%,and the mAP₅₀ is increased to 96.6%.The ADM-YOLOv11 significantly improves the detection accuracy of multi-scale targets in the dynamic scenes of tower videos.

Key words: tower video, multi-scale, YOLOv11, adaptive feature enhancement, dynamic upsampling