Tracking of Ship and Plane in Satellite Videos Using a Convolutional Regression Network with Deep Features

Abstract

Object tracking is one of the latest research topic in various applications of computer vison. Satellite videos are provided by the commercial satellite makes it possible to extend into earth observation. Moving objects are occupy the small region of pixel and they are confused with neighboring regions also with objects. Due to the resolution constraints, the nearest similar objects can rarely be distinguished by appearance features. Conventional tracking techniques in satellite videos with hand-crafted features do not give better results. Deep learning-based techniques have proven to be superior in visual tracking benchmarks, but this is not fully explored for the object tracking in the satellite videos. Regression network is used to integrate the regression model with convolutional layers. Both motion feature and appearance feature are taken from pre-trained deep neural networks VGG16, VGG19, and ResNet50. When tracker faces obscure appearance information, motion features of object are used for discrimination and complementary information to improve the performances of model. Experimental analysis on ship and plane represents the ResNet50 feature extractor gives better tracking result. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.