Motion segmentation object detection in complex aquatic scenes and its surroundings

This paper presents a novel segmentation approach for object detection in aquatic outdoor scenes such as swimming pools and its surroundings. This process is part of a video-based early drowning detection system for private swimming pools. Aquatic outdoor scenes are continuously changing due to reflections and refractions caused by water oscillation, severely affecting the results obtained by traditional segmentation schemes because water oscillation is perceived as motion detection of objects in the aquatic area. The novel segmentation technique is based in the separation of the special water environment and the rest of the scene, by means of an off-line automatically generated pool mask. The two areas are segmented with different algorithms, one for the aquatic scene and another for the surrounding area. Further, a two-step approach segmentation algorithm for the aquatic area was developed: the first segmentation step is based on the HSV invariant color model space; the second one consists in re-segmenting each region inside an expanded bounding box resulting from the first step. The second step increases segmentation quality up to 20%, as it captures parts of objects not detected before. The information extracted from the water surrounding area is also very important for object tracking and behaviour analysis purposes.