Automated Detection, Tracking Measurement, and Classification of Fish Observed by Underwater Cameras

We propose to develop a system that can automatically detect, track, measure, and classify fish seen in video collected by stationary underwater cameras. Our approach makes use of a novel method for detecting motion that can mitigate false alarms due to swaying underwater vegetation. We propose a novel technique for tracking the fish in three spatial dimensions allowing for improved knowledge of the potential target dynamics, occlusion by other fish and submerged objects, and length measurements. Our system will be capable of working with two or more cameras. Our classification module will work with multiple segmented images of fish. Classification decisions will be made using as few as one frame of data, but most decisions will be based on multiple looks at each fish, taking advantage of the fact that most video cameras collect many images per second. We propose to explore several different types of features for performing classification using a single frame of data and these features will be applied to both monochrome and RGB imagery. We will create a software tool that allows an operator to view video with overlays indicating fish in track along with their estimated length and classification.