With the increase in available computation and the miniaturization of optical sensors, autonomous surveillance systems are now capable of exploiting information gathered by cameras operating in different wave spectra, such as visible and thermal cameras. The use of these systems is particularly fitting for maritime surveillance purposes such as sea patrolling, incoming threat detection, and maritime traffic management.
In this session, you will see how this project aims to realize a system capable of supporting the decision-making process of operators involved in maritime applications. We will show the workflow from the initial design to the first system prototype deployment. We will also cover the acquisition and preprocessing of video flows coming from multispectral sensors, the transfer learning application on YOLOv2 for the detection of vessels, and the integration of the code generated by GPU Coder™ in a Visual Studio project.
The built-in functions provided by MATLAB® toolboxes for deep learning and image processing applications and the continuous support from our MathWorks worldwide friends allow us to rapidly test our ideas and, with the coder, the resulting algorithms can be translated and implemented in our projects with virtually zero effort. This has been fundamental in completing the research project in the tight 12-month deadline and having enough time to manage environment characteristics. The maritime environment, particularly the open sea, represents a challenging scenario for autonomous surveillance systems; cameras mounted on vessels are subjected to wave motion, continuous background changes, and an almost total lack of reference points. During the final tests, the system was able to correctly detect and track the selected vessel type after video flow stabilization, both in the open sea and near coast environment.
