Structural Consistency Based Multiobject Detection for Satellite Video and Benchmark With Real Scene Simulation Dataset

The accurate detection of moving objects remains a challenging task due to complex scenes and interference from light variation and background movement. Moreover, there is an urgent need for a dataset based on real scene simulation that can be used to evaluate the effectiveness of object detection methods. In this paper, we propose a satellite video moving object detection method based on Structural consistency and Spatio-Temporal bidirectional Matching (SSTM), and introduce a Large-range and Multi-class moving Object Detection dataset (LMOD) for satellite video. The SSTM first extracts moving objects by constructing a structural consistency model and conducting the weighted fusion of motion features. A Spatio-Temporal Bidirectional Matching module (STBM) is used to integrate the historical and future information, and improve the uniformity of object detection. LMOD contains the frames ranging from 1,500×1,160 to 4,000×2,000 in size, with a spatial resolution of approximately 1 meter. It labels multi-class moving objects, with total of 480,332 labeled instances. The proposed method is experimented on the proposed LMOD dataset and compared with existing state-of-the-art methods, which demonstrates its effectiveness and superiority. LMOD dataset is available at https://github.com/RS-Devotee/LMOD.