Moving Object Detection and Tracking for Satellite Video Surveillance

Download files
Access & Terms of Use
open access
Embargoed until 2022-02-16
Copyright: Zhang, Junpeng
Satellite remote sensing videos have become an emerging valuable data source for city-scale surveillance from space. Moving Object Detection (MOD) and Multiple Object Tracking (MOT) methods serve as the stone-steps for the related applications. However, they are challenged by the low spatial resolution and low signal and noise ratio. This thesis aims to meet the challenges and develop effective MOD and MOT techniques for satellite remote sensing videos. For MOD, an Extended Low-rank and Structured Sparse Decomposition (E-LSD) model is proposed to suppress the effect of random noises. E-LSD models moving objects by a sparse foreground matrix, and the structured sparse regularization is imposed on it for exploring the spatial priors on moving objects. This alleviates the false alarm problem caused by noises on satellite remote sensing videos. To promote online processing, an Online Low-rank and Structured Sparse Decomposition (O-LSD) is developed. O-LSD reformulates the E-LSD problem that combines all frames in a video to a sequence of frame-wise separable counterparts by adopting the matrix factorization approximation and stochastic optimization techniques. In this way, extracting moving objects on a frame relies only on the information from current frame and its predecessor frames. In addition to random noises, the local misalignment caused by the motion of satellite platform is another primary source of false alarms in MOD on satellite remote sensing videos. To separate moving objects from it, a Moving-Confidence-assisted Matrix Decomposition (MCMD) model is developed by integrating motion information on moving objects into the foreground regularization. This is an improvement to E-LSD model in suppressing the effect of moving satellite platform in MOD. For MOT, an Incremental Successive Shortest Path (ISSP) tracker is developed. It defines a Maximum A-Posterior (MAP) problem for selecting and linking an optimal set of detection observations for trajectory formation. By fusing the information along the extracted trajectories, the proposed ISSP tracker reduces the fragmentations in the generated trajectories while automatically discarding the false alarms, which makes it more adaptive to the scenarios with inadequate accuracy. The videos captured by SkySat and Jilin-1 satellites were utilized for the testing and evaluation. The experimental results presented in this thesis demonstrate the proposed solutions perform well, which confirms the feasibility of applying satellite remote sensing videos for reliable MOD and MOT.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Zhang, Junpeng
Xiuping, Jia
Jiankun, Hu
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
Resource Type
Degree Type
PhD Doctorate
UNSW Faculty
download public version.pdf 36.47 MB Adobe Portable Document Format
Related dataset(s)