Kernel density estimation and correntropy based background modeling and camera model parameter estimation for underwater video object detection

Abstract

Underwater video object detection is challenging because of the complex background and the movement of the camera. In order to address this, we propose a novel scheme of simultaneously estimating the camera model parameters and detecting the object. The object detection phase includes background modeling and its learning. Background is modeled by the proposed Spatial Kernel Density Estimation (SKDE) model and the model learning happens in the SKDE feature space. Background modeling and its learning is pixel based approach. The model histograms learn the new pixel through its histogram representation. Our learning and classification strategy is different from the Heikkila et al. [17] in the context of similarity measure. We have proposed the correntropy based similarity measure that is used for model learning and pixel classification. The camera model parameters are estimated by 2D optimization method where we have used the corner features of an object at subpixel accuracy level. These subpixel level features are used in the proposed pipelining framework for model parameters esti-

HLASwin-T-ACoat-Net Based Underwater Object Detection

“…Panda and Nanda [15] introduced the SKDE model, which utilizes a method of spatial kernel density estimation for learning within the SKDE feature space. Creating a model of the surroundings and gaining an understanding of it largely involves the use of individual pixels in an analytical approach.…”