Object localization in dense stereoscopic reconstruction

Rad razmatra postupak lokalizacije objekata iz oblaka točaka dobivenog stereoskopskom rekonstrukcijom. Rad se nastavlja na diplomski rad Viktora Brauta koji je razvio računalni program podešen za detekciju objekata iz para rektificiranih slika snimljenih stereo sustavom montiranim na krov vozila. U okviru rada korištena je metoda poluglobalnog podudaranja radi dobivanja gušće disparitetne mape u odnosu na lokalne metode. Rad programa je evaluiran na KITTI object ispitnom skupu. Rezultati su prikazani u obliku krivulja koje pokazuju ovisnost odziva i preciznosti o traženom minimalnom preklapanju i u obliku krivulja koje prikazuju međusobnu ovisnost preciznosti i odziva.The thesis considers the method of object localization in a point cloud. This thesis uses software for object localization developed by Viktor Braut and continues experiments described in his master thesis. This software is specifically tuned for the object detection in the rectified image sequences recorded by a pair of calibrated cameras mounted on the vehicle. The stereo matching method used for disparity map computation is Semi global block matching because it gives a better estimation of scene than local methods. Detection program was evaluated on KITTI object detection benchmark and the graphs presenting results are shown. First type of graph shows recall and precision values as a function of minimal overlapping required and the second type of graph shows precision recall curve