A neural network for trench detection based on Lidar elevation model

V nalogi naslovimo problem detekcije strelskih jarkov na digitalnem modelu višin, ki je pridobljen s sistemom Lidar. Detekcija je izvedena s pomočjo tehnologije segmentacijskih konvolucijskih nevronskih mrež, ki v zadnjih letih krojijo sam vrh pri reševanju problemov detekcije objektov ter segmentacije slik. Uspešna detekcija jarkov ima tudi zgodovinski pomen, saj avtomatske metode detekcije strelskih jarkov še niso bile preizkušene, prav tako pa sistem Lidar omogoča doslej nepredstavljivo natančno analizo terena. V nalogi predlagamo algoritem, ki temelji na arhitekturi U-net in vsebuje postopke pred procesiranja ter naknadnega procesiranja slik, saj zaradi narave problema, ki ga naslavljamo ni potrebno, da se detekcija izvaja v realnem času. Rezultate predlagane metode (Fr13) primerjamo z dvema modificiranima različicama (Fr9 ter Canny) ter sorodno metodo Edge. Primerjavo izvedemo glede na meri F1 ter MCC in pokažemo, da odvisno od tipa območja predlagane metode dosegajo od 10% do 30% boljše rezultate kakor sorodna metoda. V sklopu dela primerjamo tudi rezultate metode Fr13 ter Fr9 ter pokažemo vpliv različnega načina generacije ter perturbacije učne množice v primeru da imamo močno neuravnotežen podatkovni niz.In the thesis we address the problem of infantry trench detection in terrain images obtained with the Lidar system. Detection is performed using segmentation convolutional neural network technology, which is currently outperforming other methods when it comes to solving problems which require object detection and image segmentation. Successful detection has historical meaning as well as the automatic detection methods have not yet been used to address this problem. In addition the Lidar system is now offering a view of terrain with unprecedented precision. We present an algorithm based on the U-net architecture together with image preprocessing and post processing steps, as due to the nature of the problem the detection does not need to run in real time. We compare the results of our method (Fr13) with two modified approaches (Fr9 and Canny) and a related method - Edge. Comparison is performed using the F1 and MCC measures where our method outperforms the Edge method by 10% to 30%. Based on the different results achieved by methods Fr13 and Fr9 we present and discuss the implications different methods of learning set generation and image augmentation have on the learning process of neural networks, especially if original data is heavily unbalanced