Exploring Transformers for Intruder Detection in Complex Maritime Environment

The vast expanse of the coastline makes it difficult and expensive to deploy resources for monitoring it for safety from intruders or illegal activities. The advent of very sophisticated cameras and the myriad of object detection techniques applied to the surveillance photos and videos, provides new methods for automation of monitoring. In this paper we present a study on evaluating the various state-of-the-art object detection methods for identifying marine vessels for intruder detection. Particularly, a comparison of anchor-based, anchor-free and transformer-based object detection techniques is presented for intruder detection. Analysis on the suitability of transformer-based methods for intruder detection is also presented, with experiments performed on a combined marine vessels dataset. Our experiments show that CenterNet, an anchor-free, one-stage technique is still the fastest detection method and suited for online surveillance. Whereas the high accuracy of Transformer based methods, such as DETR, work best for offline video based surveillance.