High-level Event Detection in Underwater Video First Year Transfer Report

Video cameras are currently being employed by scientists in underwater surveys of the seabed, for various marine studies. In some instances, the duration of these surveys are quite long, which results in a mammoth of visual data that scientists have the overwhelming task of manually analyzing. It is often the case that only certain items within these recorded sequences are of scientific interest. The aim of this thesis is to try and reduce the time spent in performing this manual analysis process, by automatically identifying specific items of scientific relevance from these survey videos. Two specific items would be targeted for identification, major changes in the seabed and estimating the population of the Norway lobster, Nephrops norvegicus. Previous work involving seabed monitoring was done by Lebart et. al. in 2003, where large variations in the colour and textural features along the video sequence, were used to indicate occurances of major changes in seabed type. We plan to investigate two methods for identifying this particular scientific item of interest. The first method, would be based on detecting specific fluctuations in the three seabed constituents i.e. sand, mud and gravel. Low level features of colour, texture, shape and motion would be utilized for identifying these three constituent objects in the respective frames. The proposed framework for this algorithm is presented in this report. The second method that we would be investigating for detecting major changes in the seabed type is that of Lebart et. al. in 2003. We plan to improve this algorithm by making it robust to a number of problems observed in our test videos such as camera flashlight frames, snow frames, above water sections in the video, difficult lighting conditions, marine snow etc. From this long list of problems, we have developed and tested algorithms for detecting the camera flashlight frames, snow frames and above water regions, to date. These algorithms and their respective results obtained are presented along with preliminary research and proposed identification algorithms for the remaining problematic events. Preliminary content analysis of the underwater videos involving colour, texture and motion was also performed. With regards to the second item of interest i.e. estimating Nephrops population, we present some initial analysis of the problem, along with a proposed framework for accomplishing this task.