Using background model and shallow convolutional neural network to detect moving vehicles from satellite videos

In recent years, the Earth-observing satellites have obtained the ability to capture city-scale videos, which enable potential vehicle monitoring. Because of the broad field-of-view, the moving vehicles in satellite videos are very small, making it difficult to differentiate true objects from noise. This paper proposes a terse framework that can effectively suppress false targets while keeping a high detection ratio. The framework first applies the K-nearest neighbor (KNN) background subtraction model to produce preliminary detection results at high recall but with low accuracy, and then uses a shallow convolutional neural network (CNN) to eliminate false targets, increasing the detection accuracy. The experiments and evaluations demonstrate that our method can largely improve the accuracy at the expense of a slight reduction of recall.