Detekcija objektov na vodni gladini na podlagi zaznavanja anomalij

KALŠAN, DAMJAN (author), Kristan, Matej (mentor)

Zaznavanje ovir v avtonomnih sistemih je ključnega pomena za primeren odziv na nepredvidljive okoliščine. Na pomorski domeni so se v preteklosti raziskovalci detekcije lotili z uporabo različnih senzorjev in pogosto omenili smiselnost uporabe naprednih tehnik računalniškega vida. Zaradi neskončnega števila možnih ovir v tej nalogi problem obravnavamo v kontekstu detekcije anomalij. Tovrstne izbrane klasične metode delujejo na osnovi arhitekture koder-dekoder in se poslužujejo polnadzorovanega učenja nad slikami brez anomalij. V diplomski nalogi jih najprej primerno nadgradimo tako, da so se sposobne učiti tudi iz slik z anomalijami in nato njihovo delovanje preverimo na pomorski domeni. Za ta namen predlagamo dve prilagojeni cenilni funkciji, ki strmita k maksimizaciji rekonstrukcijske napake anomalij in minimizaciji napake morja. Predlagamo konvolucijsko nevronsko mrežo, ki na podlagi napake rekonstrukcije napoveduje segmentacijsko masko in v okviru nje primerjamo predlagani cenilni funkciji na štirih podatkovnih zbirkah. Predlagane metode se lahko uporabljajo tako na razburkani, kot tudi mirni vodni površini, vendar so zanje značilne pogoste fantomske detekcije ovir, ki za sedaj omejujejo njihovo uporabo v avtonomnem sistemu. Kljub temu metode delujejo neodvisno od bližine obale ali velikih objektov.

• Language: Slovenian
• Keywords: avtonomna plovila, konvolucijske nevronske mreže, detekcija anomalij, semantična segmentacija, strojno učenje
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2019
• PID: 20.500.12556/RUL-111811
• COBISS.SI-ID: 1538422211
• Publication date in RUL: 14.10.2019

Secondary language

• Language: English
• Title: Object detection on water surface by anomaly detection

Abstract

Obstacle detection in autonomous systems is of key importance for an appropriate response to unpredictable circumstances. A great portion of object detection research in the maritime domain so far has employed different sensors and concluded that there is a need for advanced computer vision techniques. Due to an infinite number of potential obstacles we focused on reconstruction based anomaly detection methods which utilize autoencoders for semi-supervised learning on nonanomalous images. We have assessed these methods in the maritime domain and upgraded them in a way that they are also able to learn from anomalous images. In this paper we propose two modified loss functions which strive to maximize the reconstruction error of anomalies and minimize the error introduced by the sea. We also propose a convolutional neural network which predicts a segmentation mask based on the reconstruction error using the aforementioned loss functions and test it on four different datasets. The proposed methods can be used in the case of an agitated as well as the calm sea, however due to a high number of false detections they are not yet fit for use in an autonomous system. Nonetheless the methods work in the proximity of the shore and large objects.

• Keywords: autonomous surface vehicles, convolutional neural networks, anomaly detection, semantic segmentation, machine learning