Detekcija in klasifikacija objektov v vodnem okolju s pomočjo konvolucijskih nevronskih mrež

Ambrožič, Nejc (author), Kristan, Matej (mentor)

Problematika, ki jo naslavljamo v magistrskem delu, je detekcija objektov s konvolucijskimi nevronskimi mrežami v vodnem okolju za detekcijo nevarnih ovir v realnem času. Zanesljiva in hitra detekcija ovir je osnovni problem avtonomne vožnje. Konvolucijske nevronske mreže so pogosto uporabljene v avtonomnih avtomobilih, v vodnem okolju pa še niso bile temeljito preizkušene. V magistrskem delu analiziramo dve izmed najnovejših konvolucijskih nevronskih mrež za detekcijo in klasifikacijo objektov: YOLO in BlitzNet. Predlagamo modificirano konvolucijsko nevronsko mrežo YoloW in novo podatkovno zbirko za detekcijo objektov v vodnem okolju WODD. Podatkovna zbirka vsebuje 19691 anotiranih ovir, ki se pojavijo v 12168 slikah. Predlagamo tudi prilagojen postopek učenja v podatkovni zbirki z negotovimi učnimi primeri, ki je primeren za učenje konvolucijskih nevronskih mrež na realnih podatkovnih zbirkah. V delu evalviramo delovanje predstavljenih konvolucijskih nevronskih mrež na predlagani podatkovni zbirki. S povečevanjem učne množice se natančnost vseh predstavljenih mrež izboljšuje. Po učenju na celotni testni množici podatkovne zbirke doseže mreža BlitzNet povprečno natančnost 89.68%, YOLO 96.78%, medtem ko naša mreža YoloW doseže 97.72%. Mreža YoloW deluje v realnem času in je sposobna detekcije ovir v povprečno 30.12 slikah na sekundo.

• Language: Slovenian
• Keywords: računalniški vid, strojno učenje, globoko učenje, konvolucijske nevronske mreže, detekcija objektov
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-105409
• Publication date in RUL: 26.11.2018

Secondary language

• Language: English
• Title: Object detection and classification in aquatic environment using convolutional neural networks

Abstract

We address the problem of real-time floating obstacle detection in aquatic environments with convolutional neural networks. Reliable and fast detection of obstacles is crucial for autonomous driving. Convolutional neural networks are often used in autonomous cars but have not yet been thoroughly tested in the aquatic environment. For this purpose, we analyze two of the latest convolutional neural networks for object detection and classification: YOLO and BlitzNet. We propose a modified convolutional neural network for obstacle detection YoloW and a new dataset for object detection in the aquatic environment. The dataset contains 19691 annotated obstacles appearing in 12168 images. We propose a customized learning process from uncertain training examples, which is suitable for training convolutional neural networks on real world datasets. We evaluate the performance of the presented convolutional neural networks on the proposed dataset. By increasing the number of training examples, the accuracy of all models is improved. After training on the entire training set of our dataset, BlitzNet achieves an average accuracy of 89.68%, YOLO 96.78%, while our model YoloW achieves an average accuracy of 97.72%. The proposed YoloW works in real-time and is capable of obstacle detection at 30.12 images per second on average.

• Keywords: computer vision, machine learning, deep learning, convolutional neural networks, object detection