Underwater Target Tracking via 3D Convolutional Networks

The task of underwater target tracking is one of the most important challenges in recent upsurge of smart aquaculture, especially in the application of AI-driven cage culture. However, tracking often requires short computational time and draws little attention from researchers in the field of deep learning. A convolutional network tracker (CNT) was proposed [1], which uses 2D features of local structure and internal geometric layout information between the target candidates in adjacent frames to address the tracking tasks without pre-training. In [14], an improved version of CNT (called Fast-CNT) was proposed for performing underwater multi-target tracking. This paper further proposes a 3D version of CNT (called 3D-CNT) characterized by extracting spatiotemporal features between successive frames to make the target (e.g. fish) template more robust in tracking. Experimental results show that with these temporal features, 3D-CNT outperforms the Fast-CNT in tracking moving fish.