Detecting traversable terrain by stereoscopic computer vision

Rad se bavi tehnikama detekcije rubnika kao dijelom sustava za pomoć u vožnji. Ulazni podaci u implementaciji su slike pribavljene stereo parom kamera. Implementirane su metode rekonstrukcije trodimenzionalne scene iz slika u strukturu oblaka točaka. Ta se struktura projicira u kartu visina. Nakon sto je konstruirana karta visina, moguće je koristiti dvije metode detekcije. Jednostavna metoda detektira rubove na karti visina, te među rubovima traži pravce koje iscrtava kao rubnike. Druga metoda uključuje estimiranje parametara modela prometnice, tj. definiciju površina ceste i pločnika i krivulje rubnika kao granice među njima. Rezultate dobivamo nakon projekcije karte visina u uvjetno slučajno polje i klasifikacije. Rubnik se iscrtava na originalnim slikama. Prezentirani su rezultati eksperimenata na ispitnim skupovima.This thesis is concerned with techniques of curb detection, as a part of a advanced driver assistance system. Input to the algorithm implementation are pictures acquired by a stereo rig on a vehicle. Thesis includes implementation of methods for extracting the three-dimensional information from the pictures into a point cloud structure. Point cloud is then projected into a digital elevation map. After building the elevation map, it is possible to make use of any of two methods for curb detection. The simpler method detects edges on the digital elevation map and locates lines in edge map. Those lines are projected back to the image and represent curbs. The other method includes parameter estimation for the environment model, i.e. defining the surfaces of the road and sidewalk and defining the curb curve as an edge between the two surfaces. Results are acquired after projecting the elevation map into a conditional random field and using an inference algorithm for classification. The drawings of the curb are then added to the input images. Results of testing the algorithms are also presented