Adaptive Feature Fusion Based on Double Head Framework for Small Object Detection in Highway Scenarios

Object detection has achieved unprecedented advances in recent years as a major vision mission. However, small object detection still faces many challenges, such as the lack of pixels and background disturbances. Therefore, in this paper, we introduce adaptive feature fusion based on a double-head framework (AFFD) for small object detection on highways. The AFFD has two core modules, including a layers attention (LAT) and a double head (DH). Firstly, the LAT is designed to fuse the output of the feature pyramid and enhance the representation of small objects. Secondly, inspired by the properties of fully connected and convolution, we propose the DH to obtain the precise classification score and regression bounding box. To verify the performance of our method, we establish a dataset, named Small Object Detection on Highways (HSOD), with a total of 931 images and 1,458 instances. The HSOD dataset has issues with tiny object detection, including imbalanced item samples, poor geographic distribution, and pixel deficiency. It reflects real-world challenges and is thus well-suited for tiny object detection. In addition, we introduce a data augmentation method to supplement and refine the dataset. Finally, experiments are conducted on our proposed HSOD dataset, AI-TOD, Marine Ship, MSCOCO 2017 and SODA-D. Experimental results show that our AFFD has achieved excellent performance compared with state-of-the-art methods. The data is released athttps://github.com/HSOD.