Towards Maritime Videosurveillance Using 4K Videos

摘要

This paper develops a novel approach to automatic maritime target recognition in the framework of near real-time maritime video-surveillance using super-resolved (i.e.; 4K) videos captured either with a static or with a moving video camera. The challenge of achieving a robust 4K video-based surveillance system is twofold. Firstly, the 4K video resolution (3840 × 2160 px.) constrains considerably the amount of video-processing for meeting the near real-time requirement. Secondly, maritime environment is very dynamic and highly unpredictable, thereby, rendering target extraction a difficult task. Therefore, the proposed approach attempts to leverage both temporal and spatial video information for achieving fast and accurate target extraction. In fact, since, the object rigidity assumption is implemented parsimoniously, i.e.; at key video locations, its real-time implementation, first, enables to quickly extract potential (sparse) target locations. Furthermore, we have shown, experimentally using many maritime videos, that maritime targets generally exhibit richer textural variations than dynamic background at different scales. Thus, secondly, a still image based multi-scale texture discrimination algorithm carried out around previously extracted key video locations allows to achieve final target extraction. An experimental study we have conducted both using our own maritime video datasets and publicly available video datasets have demonstrated the feasibility of the proposed approach.