Motion trajectory estimation of salmon using stereo vision

Abstract

A main concern for the aquaculture industry is the fish behaviour and welfare. Motion trajectory analysis of salmon at aquaculture farming sites with respect to certain aquaculture operations aims to provide information about the behaviour and possibly stress level of the farmed salmon and may help to generate a general welfare indicator index. Towards this aim we present an innovative computer vision and machine learning based approach for motion trajectory estimation of salmon. Video footage was recorded with a stereo camera setup. Deep learning based object detection was performed to detect particular features. We focused on tracking the fish eyes and heads as a reliable indicators of the fish's position. Feature matching and subsequent 3D reconstruction was performed to calculate the 3D position of the fish from which trajectories of the fish movement were estimated. Related experiments were conducted at an aquaculture research facility under natural lighting conditions and extracted trajectories allowed a qualitative verification. The developed method was verified using synthetic ground truth data produced with an open source computer graphics software for quantifiable performance metrics.