To address the insufficient accuracy of small object detection by unmanned aerial vehicle in complex highway scenarios

a small object detection algorithm that utilizes adaptive feature fusion and subspace attention mechanisms is proposed.This method takes the unmanned aerial vehicle onboard image sensor as the data source and innovatively employs soft pooling downsampling technology to preserve the edge features of small objects.It dynamically balances multi-scale feature fusion through learnable parameters and designs an optimized subspace attention module to enhance the representation of small object features.Experimental results demonstrate that on the BDD100K and KITTI datasets

this method significantly outperforms existing mainstream methods with mean average precision (mAP) values of 56.36% and 93.33%

respectively

which validates its effectiveness in detecting small objects in complex environments.This research provides a more accurate solution for unmanned aerial vehicle-based highway environment perception.