Real-Time Ship Monitoring Algorithm and Measurement for Single-Shot Multi-Box Detection and Navigation Based on Deep Learning

EasyChair Preprint 3075

Abstract

Man-made ship collisions have greatly affected the marine environment. Automatic identification system (AIS) is widely used. Currently, researchers use AIS-based data analysis methods to predict navigation risks, but they have not been able to solve the problem of real-time ship detection. This paper studies the real-time detection method of a ship's single-shot multi-box detector (SSD) framework with typical application scenarios. SSD is a single-order deep convolutional neural network (CNN) learning algorithm that uses a feedforward CNN to generate a set of fixed-size bounding boxes for each object from a different feature map. We evaluated a number of different feature extractors, including Faster RCNN (VGG16), YOLO, YOLOv2, YOLOv3, SSD300, SSD512, RefineDet320, RefineDet512. In 2019, we collected a detection dataset of a ship sailing video atlas. To verify the method, we identified it using various ship detection methods and compared the SSD with YOLO and RefineDet. Our results show that the method has good test results and surpasses all ship detection methods. Specifically, in terms of detection speed, our proposed method is superior to all methods and can meet the actual needs of ships when detecting ships in the surrounding waters in real time. In short, the SSD-based real-time ship detection method performs well and has the potential to improve accuracy and efficiency.

Keyphrases: Automatic Identification System, YOLO, single shot multi-box detector

@booklet{EasyChair:3075,
  author    = {Xiang Wang and Jingxian Liu and Zhao Liu and Zhi Yuan},
  title     = {Real-Time Ship Monitoring Algorithm and Measurement for Single-Shot Multi-Box Detection and Navigation Based on Deep Learning},
  howpublished = {EasyChair Preprint 3075},
  year      = {EasyChair, 2020}}