Multi-Scale Feature Fusion and Multi-Head Attention Seam Module for Occluded Shared Bike Detection

In recent years, shared bicycles have become an important way of short-distance travel in cities, but the problem of random parking needs to be efficiently managed. In the occluded scene, the existing detection methods are prone to misdetection or omission of occluded areas and small targets due to the variable target morphology caused by occlusion, the difficulty of feature extraction, and the loss of details caused by downsampling. In this paper, an improved model based on YOLOv11, which integrates multi-scale features and multi-head attention SEAM Module, is proposed to solve the problem of occluded target detection in the above-mentioned shared bicycle images. Firstly, by introducing a multi-scale feature fusion module based on deconvolution and replacing the original upsampling and splicing operations, the fusion effect of multi-scale features is enhanced, so that the model can more effectively capture semantic information at different levels. Secondly, the multi-head attention SEAM (Separated and Enhancement Attention Modul) Module is introduced before the detection head to improve the model's ability to detect occluded targets by adaptively weighting important features and suppressing redundant information. Experimental results demonstrate significant improvements in recall (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$+5.5 \%$</tex>), and mAP50 (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$+3.3 \%$</tex>), particularly in complex occlusion scenarios, validating the effectiveness of both modules.