Assessing Deep-learning Methods for Object Detection at Sea from LWIR Images. Issue 21 (2019)

• Record Type: Journal Article
• Authors: Schöller, Frederik E.T.; Plenge-Feidenhans'l, Martin K.; Stets, Jonathan D.; Blanke, Mogens
• Is Part Of: IFAC-PapersOnLine. Volume 52:Issue 21(2019)
• Page Start: 64
• Page End: 71
• Publication Date: 2019
• Subjects: Object Detection -- Autonomous marine crafts -- Navigation -- Long-wave Infra-red -- Detection at sea -- Autonomous Ship; Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2019.12.284
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: This paper assesses the performance of three convolutional neural networks for object detection at sea using Long Wavelength Infrared (LWIR) images in the 8- 14µm range. Capturing images from ferries and annotating 20k images, fine-tuning is done of three state of art deep neural networks: RetinaNet, YOLO and Faster R-CNN. Targeting on vessels and buoys as two main classes of interest for navigation, performance is quantified by the cardinality of true and false positives and negatives in a random validation set. Calculating precision and recall as functions of tuning parameters for the three classifiers, noticeable differences are found between the three networks when used for LWIR image object classification at sea. The results lead to conclusions on imaging requirements when classification is used to support navigation.