Robust visual place recognition using deep representations and sequence-based image matching ; Robusno vizualno prepoznavanje mjesta uporabom dubokih reprezentacija i uparivanja sljedova slika

The doctoral thesis examines the problem of visual place recognition. During visual place recognition, negative impacts such as different environmental conditions, different viewpoints of the camera and perceptual aliasing, affect the system’s performances. Therefore, the thesis proposes a novel sequence-based place matching method. Such a place matching is more robust than the naïve approach because data is sequential and captured during the vehicle’s traversal around a route. Also, the thesis proposes a robust visual representation for a place image by using fine-tuned deep convolutional neural networks. Feature maps extracted with a convolutional neural network are further refined with the mutual information-based feature selection technique. Finally, to not be an end in itself, the implemented system is additionally adapted and used in the context of the famous robotic problem – simultaneous localization and mapping – for the purpose of loop closing detection where even better performances are obtained. Scientific contributions of the thesis are: Sequence-based visual place matching method that uses directed acyclic graphs and single source shortest path. Method for robust visual place recognition with deep representations that uses softmax regression and mutual information-based feature selection. Procedure for adapting sequence-based visual place recognition method for loop closing in simultaneous localization and mapping algorithms. ; Doktorski se rad odnosi na problem vizualnog prepoznavanja mjesta. Pri vizualnom prepoznavanju mjesta, negativni čimbenici kao što su različiti okolišni uvjeti, različita gledišta kamere te vizualna sličnost različitih mjesta, utječu na performanse sustava. Stoga ovaj doktorski rad predstavlja novu metodu prepoznavanja mjesta uparivanjem sljedova slika. Takvo je uparivanje robusnije od uobičajenog pristupa jer su podaci nastali pri obilasku vozila oko nekog područja sekvencijalne prirode. Također, doktorski rad predstavlja robusnu vizualnu reprezentaciju slike mjesta ...