Analiza temporalnih segmentacijskih metod za detekcijo ovir na domeni avtonomnih plovil

ŠKODNIK, LUKA (author), Kristan, Matej (mentor), Žust, Lojze (co-mentor)

Semantična segmentacija je uveljavljen pristop reševanja problema navigacije avtonomnih vozil. Okolica avtonomnih plovil je zelo specifična in se razlikuje od splošne okolice avtonomnih vozil. Zaradi tega so potrebne podatkovne zbirke, metode za pridobitev semantične segmentacije in načini evalvacije posebej zasnovani za vodno domeno. V sklopu naloge anotiramo manjšo podatkovno zbirko z raznolikimi prizori vodne okolice, ki je vkjučena v večjo podatkovno zbirko, ki je uporabljena za učenje in evalvacijo. Za izboljšanje semantične segmentacije videoposnetkov nekatere metode uporabljajo časovno informacijo v obliki zaporednih videosličic. Dve taki metodi učimo s podatki iz vodnega okolja in ju primerjamo z metodo zasnovano posebej za pridobitev segmentacije v vodnem okolju za navigacijo avtonomnih plovil. Zaradi premajhne splošnosti uporabljenega evalvatorja, priredimo način evalvacije in dodamo nekaj novih mer. S temi merami na splošno ocenimo segmentacijo, detekcijo dinamičnih ovir in zaznan vodni rob ter podamo interpretacijo rezultatov. Ugotovimo, da ena izmed metod, ki uporablja časovno informacijo v splošnem daje boljše rezultate od druge, vendar je druga bolj robustna pri zaznavanju težkih primerov na vodni gladini (npr. odsevi). Prav tako ugotovimo, da ti dve metodi samo z uporabo časovne informacije ne pridobita dovolj, da bi podali boljše rezultate od metode, ki je zasnovana specifično za vodno okolje. Zadnja ugotovitev je, da na rezultate zelo vplivajo tudi učni podatki. Če metode učimo na podatkih splošnega vodnega okolja, dobimo veliko boljše rezultate, kot če metode učimo samo na podatkih iz morskega okolja.

• Language: Slovenian
• Keywords: semantična segmentacija, detekcija ovir, avtonomna plovila
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science ; FMF - Faculty of Mathematics and Physics
• Year: 2022
• PID: 20.500.12556/RUL-139798
• COBISS.SI-ID: 121660419
• Publication date in RUL: 07.09.2022

Secondary language

• Language: English
• Title: Analysis of temporal segmentation methods for obstacle detection on the autonomous vessels domain

Abstract

Semantic segmentation is an established approach for obstacle detection in autonomous ground vehicles. The environment of unmanned surface vehicles is very specific and differs from the general environment of autonomous ground vehicles. Because of this we need datasets, methods for semantic segmentation and ways of evaluating specifically designed for the water environment. As a part of this thesis we annotate a small dataset with diverse scenes in a water environment, which is included in a bigger dataset used in our evaluation. Some methods for video semantic segmentation use temporal information in the form of consecutive frames to improve their results. We train two such methods with data from a diverse water environment and compare them to a method designed specifically for semantic segmentation of a water environment for navigation of unmanned surface vehicles. Because of the lack of generality in the used evaluator, we adapt the evaluation and add a few new measures. With this measures we evaluate the segmentation, the detection of dynamic obstacles and the detected water-edge. Lastly we give our interpretation of the results. Our first finding is that one of the methods which uses temporal information in general gives better results than the second one, however the second one is more robust in presence of difficult cases on the water surface (e.g. reflections). We also find out that these two methods don’t gain enough with the use of temporal information to surpass the method that is specifically designed for a water environment. Our last finding is that the training data greatly impacts the results. Training the methods on data from a general water environment gives much better results than training the same methods on data only from a sea environment.

• Keywords: semantic segmentation, obstacle detection, unmanned surface vehicles