Re-identification of Vessels with Convolutional Neural Networks
Amir Ghahremani, Yitian Kong, Egor Bondarev, Peter H. N. de With · 2019
In order to perform a reliable vessel behavior analysis for maritime surveillance, re-identification of previously detected vessels, passing through new camera locations, is of vital importance. However, challenging outdoor conditions of the maritime environment heavily restrict the application of conventional methods. Additionally, vessels are large objects and capturing a vessel from different viewpoints may provide entirely different visual appearances. To address these challenges, this paper proposes an Identity Oriented Re-identification network (IORnet) for the re-identification of vessels. This CNN-based approach incorporates the triplet loss method combined with a new loss function, which leads to improved vessel re-identification. Experimental results on our real-world evaluation dataset reveal that the proposed method achieves 81.5% and 91.2% on mAP and Rank1 scores, respectively. As an additional contribution, we also provide our annotated vessel re-identification dataset to the open public access.