Maritime surveillance is an important research topic, considering that it is vital to keep the maritime environment safe against dangers like terroristic attacks and illegal activities such as out-of-region-fishery, human and drugs traffic. In a general maritime surveillance system, besides radar systems in large harbors, visual cameras are deployed along the shorelines which capture the videos of the maritime environment. As the main task for such a system, vessels moving in the maritime region should be detected and their behavior analyzed. However, while aiming at finding maritime vessels, the surveillance cameras also capture multiple non-relevant regions in the scene, containing lots of moving objects, such as humans, trees, cars and clouds. As a result, suboptimal ship detectors may extract a lot of non-relevant objects as maritime vessels. Consequently, the surveillance system is supposed to segment the maritime region and then only analyze the objects located in the extracted water region.
The project focuses on research and development of innovative method for automatically extract the water regions in arbitrary scenes. We have already developed a spatiotemporally-oriented energy (SOE) feature based method to extract the water region. In our method, the mean shift algorithm smooths the SOE feature space and clusters the scene into coherent regions. Then a raster-order based labeling algorithm (ROLA) is applied to assign the same labels to clusters with corresponding properties. Finally, water regions are identified against water identification thresholds.
Despite reasonable results published in two papers, still there are interesting challenges in this work:
- Clustering block fails to separate regions with similar appearance (e.g. the water and sky in a dark or foggy environment)
- This method expects dynamic behavior from water. However, water pixels present different behaviors
- Dynamic background (wind-blown vegetation, full of car bridges, etc.) behavior.
In order to propose a robust water region extraction method, two solutions can be investigated. First, improving our clustering block by utilizing other grouping algorithm and define better criteria; second, convolutional neural networks.