Reconstruction of 3D Information about Passing Vehicles

Petr DobeŇ°

Supervisor(s): Adam Herout

Brno University of Technology

Abstract: This paper reports experiments performed during the work on my master's thesis which focuses on 3D reconstruction of vehicles passing in front of a traffic surveillance camera. Calibration process of surveillance camera is first introduced and the relation of automatic calibration with 3D information about observed traffic is described. Afterwards, a set of experiments with feature matching and Structure from Motion algorithm are presented and their results on images of passing vehicles are examined. Modifications to correspondence search stage of Structure from Motion pipeline are then proposed. Most importantly, instead of using SIFT features, DeepMatching algorithm (originally devised to find quasi-dense point matches in optical flow calculation) is used to obtain point correspondences for subsequent reconstruction phase. As a result of implemented modifications, the overall completeness of reconstructed point cloud model of passing vehicle has improved significantly.
Keywords: 3D Reconstruction, Computer Vision, Image Processing
Full text:
Year: 2017