Using motion for improved maritime obstacle detection

Metode za detekcijo ovir na podlagi slike so ključnega pomena za varno plovbo avtonomnih robotskih plovil. Trenutno najboljša metoda WaSR izvaja detekcijo ovir preko semantične segmentacije trenutne slike z upoštevanjem inercijskega senzorja. Pri tem pa ima, tako kot sorodne metode, težave pri majhnih objektih, odbleskih in odsevih. V nalogi predlagamo upoštevanje gibanja za razreševanje vizualne negotovosti v segmentaciji trenutne slike. Predlagamo postopek, ki s poravnavo preteklih slik obogati izgled slikovnega elementa v trenutni sliki s preteklimi izgledi. Obogatena slika se nato segmentira z minimalno modificirano arhitekturo WaSR. Korespondence trenutnega slikovnega elementa s preteklimi se izračunajo s trenutno eno najuspešnejših metod izračuna optičnega toka RAFT. Predstavljenih je več variant zapisa obogatene vizualne informacije. Rezultati evalvacije pokažejo podobno F_1 mero WaSR_Mu verzije modela v primerjavi z originalnim, ko upoštevamo vse ovire, ter enako dobro detekcijo roba morja. Glavna pridobitev pa je 12% izboljšava na detekcijah znotraj nevarnega območja. Vzrok za boljše rezultate je predvsem v manjšem številu lažno pozitivnih detekcij ob omenjenih težjih pogojih.Image-based obstacle detection methods are crucial for the safe navigation of autonomous robotic vessels. Current state-of-the-art method WaSR performs obstacle detection with semantic segmentation of current image while using information from inertial sensor. However, like related methods, it has problems with small objects, glare and reflections. In this thesis, we propose a method which takes into account movement to resolve visual uncertainty in the segmentation of current image. We propose a process that enriches the appearance of the image element in the current image with past appearances by aligning past images. The enriched image is then segmented with a minimally modified WaSR model. The correspondences between current image elements and ones in previous images are calculated using currently one of the most successful methods of calculating optical flow RAFT. Several variants of using enriched visual information are presented. The results of the evaluations show similar F_1 measure of the WaSR_Mu model compared to the original one while taking into account all obstacles and comparable detections of the water edge. The main achievement is 12% imporvement in detecting obsticles within the danger zone mainly due to smaller number of false-positive detections under mentioned difficult conditions