Tracklet-based vessel re-identification for multicamera vessel-speed enforcement

In crowded waterways, maritime traffic is bound to speed regulations for safety reasons. Although several speed measurement techniques exist for road traffic, such systems are not available for vessels. This paper proposes a new approach for tracklet-based re-identification (re-ID) as a solution for vessel-speed enforcement. For evaluation, the Vessel-reID dataset is used that we introduced in previous work [2]. The core of the tracklet re-ID approach is based on a novel
Tracklet-based Querying Procedure as a more effective alternative to the Common Querying Procedure (CQP) found in popular re-ID datasets [7, 8]. The existing procedure randomly selects a single image from the whole query-vessel
trajectory (in one camera view). This is improved by (1) detecting a set of most representative images per tracklet of a query-vessel, and by (2) raising the matching accuracy based on accumulating the gallery similarity scores for all images
in the set. In the experimental validation, we adopt two well-known person reID algorithms, TriNet [3] and MGN [6], since most re-ID literature focuses on person re-ID. Results show a significant increase in performance by applying the
tracklet-based approach instead of CQP: a gain of 5.6% and 8.1% Rank-1 for MGN and TriNet, respectively.