Detection and classification of marine vehicles

Abstract

One of the most common tasks within the computer vision field is the detection and classification of different objects. This thesis aims to deliver a software that can be deployed into real world scenarios and mange to detect and classify marine vehicles accurately. Using one of the pre-defined deep neural network models You look only once (YOLO), we managed to achieve a high performance for the detection and classification task. The training of the model took place using a specific dataset of grayscale images, which led to a model that can classify the objects with an accuracy of 68% and predict the relevant position with mean average precision (mAP) of 0.77. Moreover, the model tested into different weather conditions and achieved an accyracy of 0.85% and mAP of 0.068. In general, the YOLO model seems to be a robust detector that can be trained and deployed for detecting efficiently objects with high performance. Keywords: