Options
2018
Doctoral Thesis
Title
Change Detection in Combination with Spatial Models and its Effectiveness on Underwater Scenarios
Abstract
In the last years, cheap off-the-shelf consumer cameras became more and more available while the imaging quality of these devices increased drastically. This is not only true for normal in-air cameras but also for underwater devices. In connection with a growing interest in the sea world by science and industry, this led to the need for advanced computer vision algorithms, especially for underwater scenarios, so that the vast amount of acquired data can be automatically processed and important information extracted. However, apart from the normal difficulties of computer vision – the creation of detections that are consistent and coherent in the spatial as well as temporal domain – new problems occur in the underwater world caused by the specialties of the medium water. These difficulties are, among others, blur, color cast, Marine Snow and refraction which all complicate any computer vision task. Therefore, this thesis proposes a novel change detection approach and combines it with different especially developed spatial models, this allows an accurate and spatially coherent detection of any moving objects with a static camera in arbitrary environments. The spatial models include a novel approach based on the idea behind Ncut but also a derivative of the Markov Random Field model. To deal with the special problems of underwater imaging, different enhancement algorithms were used in combination with the change detection approach to counter the negative effects, e.g. Marine Snow removal or a learning-based deblurring. Furthermore, pre-segmentations based on the optical flow and other special adaptions were added to the change detection algorithm so that it can better handle typical underwater scenarios like a scene crowded by a whole fish swarm. To use these detections and extract the most information out of them, a blob tracking method is introduced that keeps the generality of the change detection approach but can still deal with occlusions or detections errors. Overall, the here presented novel general detection and tracking approach can deliver accurate results in almost all scenarios while dealing especially well with underwater videos. The different steps of the pipeline were tested and compared on different datasets against many state of the art algorithms and showed competitive results. For the testing on underwater videos a new special dataset had to be created – since this area was grossly neglected so far – and there the proposed change detection could outperform any existing method thanks to the newly developed adaptions.
Thesis Note
Rostock, Univ., Diss., 2018
Author(s)
Open Access
Rights
CC BY-NC-ND 4.0: Creative Commons Attribution-NonCommercial-NoDerivatives
Language
English