Fish Motion Approximation using ML-based Relative Depth Estimation and Object Tracking

Fish motion is a very important indicator to express various fish statuses in fish farming. Current methods in the literature for fish motion estimation either depend on special sensors or use velocity computation formulas to resolve swimming velocity but only from a 2D spatial perspective. The latter can not accurately represent the fish velocity in 3D space. We use multi-object tracking and a monocular depth model to extract information about fish location information from the recorded video, then use a post-processing approach that we designed to process this information and estimate fish velocity in 3D space. Our results show that our estimated fish velocity is approximate to the groundtruth velocity very well and that the errors remain acceptable for the use case of fish monitoring across all fish scenes. The results emphasize that although the multi-object tracking and mono-depth models both have inaccuracies that impact fish motion estimation, these impacts are not large enough to make the estimated velocity diverge away from the groundtruth. Thus, we believe that multi-object tracking and mono-depth models may demonstrate a cost-effective solution to adequately estimate fish motion in 3D space.