Augmenting Discriminative Correlation Filters with Stereo Blob Tracking for Long-Term Tracking of Underwater Animals

This paper presents a vision-based model-free longterm tracking algorithm to be used on-board autonomous underwater vehicles (AUVs) for long duration marine animal observation missions. During underwater tracking missions, drifting and losing track of targets after they leave the field of view are two major problems with state-of-the-art tracking algorithms. To achieve the long-term tracking goal, the proposed method gained drift resistance and target re-capturing ability by combining the merits of two mature short-term trackers: stereo blob tracking and discriminative correlation filter (DCF). In our approach, stereo blob tracking acts as complementary supervision to correct drift and to guide DCF to learn target appearances online before any tracking interruptions. The target information learned is then used to help re-capture the target after a tracking failure. In our experiments on field data, compared to running DCF alone, running the proposed augmented tracker increased average bounding box accuracy by 45% and eliminated drift-caused tracking failures. Our tracking algorithm also achieved 86% target re-capturing success.