Applying convolutional networks to underwater tracking without trainin

Abstract

Underwater target tracking is one of the most important tasks in recent upsurge of smart aquaculture, especially in the application of AI-driven cage culture. In [1] CNT was shown quite effective in the task of land-based object tracking, it is fast as no huge amount of data is required for training. However, an initial bounding box must be selected manually in CNT for tracking a single object, not to mention the fact that applications of CNT and other convolutional networks in underwater operations are rarely reported. In this paper we present an improved version of CNT (called Fast-CNT2) which is capable of performing underwater multi-target tracking. Firstly, GMM is applied to the input video for extracting regions each containing a moving target (e.g. fish); then a respective region containing a fish is identified and bounded with a box; finally, multi-target tracking is implemented with Fast-CNT2. Experimental results show that our method can successfully track multiple fish.