IMAGE-BASED OBJECT DETECTION AND TRACKING METHOD FOR SHIP NAVIGATION

摘要

Situational awareness technology can contribute to safe navigation by automatically detecting and alerting potential collisions in advance. In this study, we implement computer vision-based object detection and tracking for the situational awareness in the maritime environment. For object detection, one of the state-of-the-art detection algorithms using a convolutional neural network (CNN) is applied. On the top of the two-stage detection algorithm based on Faster R-CNN, we utilize feature pyramid network (FPN) to achieve high recall score especially for small scale obstacles within a wide field of sight on the open sea. For object tracking, we utilize conventional tracking method using a Kalman filter and Hungarian algorithm. When the new detection result of the current image frame is fed into the tracker, new tracking position is predicted in association with internal tracking information built upon up to the previous frames. This study shows the collaboration of deep learning-based object detection and conventional tracking method can feasibly perform the situational awareness in the maritime environment.