Saliency Subtraction Inspired Automated Event Detection in Underwater Environments

Abstract

Unmanned underwater exploration in unconstrained environments is a challenging problem. Analysis of the large volumes of images/videos captured by underwater stations/vehicles manually is a major bottleneck for further research. Existing computer vision methods either do not target unconstrained underwater environments or they only aim to detect static or moving entities. In this paper, we present a novel method for analyzing underwater videos and detecting events. Entry/exit of an object in scene is treated as an event independent of the other objects present therein. The method is applied on underwater videos with no prior knowledge, thus aiding in automated underwater exploration. The method is inspired by the fact that saliency of objects in the scene is invariant of the surrounding environment. The proposed method is composed of three main steps: Local Patch Saliency, Adaptive Saliency Subtraction, and event generation for analyzing underwater imagery from the videos. The method is aimed at detecting overlapping events containing man-made as well as natural objects including those containing multiple objects in the unconstrained underwater conditions. The performance of the method is evaluated on publicly available videos obtained from Ocean Networks Canada and Fish4Knowledge datasets. Ground truth for Ocean Networks Canada videos is not available; hence, a method for generating the same for varied sources is also presented. The algorithm achieves a precision of 98% for event detection with 20% misclassification rate. The results show the robustness of the method that performs even in complex and varying underwater conditions.