Fourier Feature Pyramid Network for Improved Small Object Detection in Complex Traffic Scenes

Small object detection in traffic scenes remains a challenging task due to the limited resolution and context of such objects. Existing feature pyramid networks primarily focus on feature layer interactions,often neglecting the difficulty of representing small object features within inner feature layers.To address this issue,we propose the Fourier Feature Pyramid Network (FFPN),which learns multi-domain representations of small objects within shallow feature layers to enhance multi-scale feature fusion and improve detection performance.By incorporating Fourier transform-based global information,FFPN compensates for information loss caused by convolution and pooling operations,thereby strengthening the feature representation of small objects.Our experimental results on benchmark datasets,including MS COCO, UAVDT,and UA-DETRAC,demonstrate the effectiveness of our approach.Notably,we achieve state-of-the-art performance on the UA-DETRAC dataset with an average precision (AP) of 91.50% for small objects, representing a 1.45% improvement over existing methods.Here we show that our method holds promise in applications such as autonomous driving,enabling accurate recognition of small targets like long-range traffic signs.