LGFF-Net: Airport Video Object Segmentation based on Local-Global Feature Fusion Network

In this work, we address the task semi-supervised video object segmentation for airport video surveillance and explore an effective solution to the specific challenge in the large-scale airport scene. We proposed a novel pipeline named Local-Global Feature Fusion Network (LGFF-Net), which can produce segmentation result in an end-to-end manner without any online fine-tuning. LGFF-Net consists of three main parts, including Global Encoder, Local Encoder and Joint Decoder. The global segmentation branch considers the comprehensiveness of the characteristics of the entire scene, ensuring the integrity of the segmentation results. The local segmentation branch focuses on obtaining the richer appearance features of the interest and is responsible for the accuracy of the results. After that, we comprehensively concern the completeness and accuracy of the target and fuse the features of each part through joint decoding. The whole network is not only clear and easy to train, but also robust to small objects in airport ground. Our method has been applied on the Airport Ground Video Surveillance benchmark (AGVS), and experiments show the effectiveness of our algorithm.