Restoration and clustering of overexposed background elements for vision systems in dynamic aquatic scenes

Overexposure causes a multitude of problems when videos of highly dynamic environments are used for computer vision applications. Foreground separation, feature detection and clustering operations get severely affected. One reason for this is the saturation or clipping of the color channels in cameras, under bright illumination conditions. Clustering the similar background data in a single cluster under partial clipping of pixel values in the RGB space and correcting the clipped data is of great importance. An analysis of the pixelwise distributions in stationary cameras trained on dynamic aquatic conditions under controlled environments and certain real world scenarios such as in swimming pools is presented in this paper. It was observed that the pixel distributions can be matched to cylindrical clusters, based on their statistical properties. It was also identified that partial clipping of color channels project the distributions onto these bordering regions of the RGB space where saturation occurs. A projection mechanism that collectively clusters saturated and unsaturated pixels that belong to the same background dynamics is presented. A method for correcting the pixel values based on the available temporal information has also been proposed. The results show that the method can successfully match the pixel values from backgrounds that undergo partial clipping correctly to the cluster that contains unclipped data from the same background dynamics. It updates the cluster suitably using the undistorted information. A statistically coherent restored value is also generated, which can be used for higher level processing such as the calculation of spatial and temporal frequencies, as well as for the restoration of frames with suitable post processing.