Estimating OD matrices at intersections in airborne video - a pilot study

Viktor Braut, Mateja Čuljak, Vedran Vukotić, Siniša Šegvić, Marko Ševrović, Hrvoje Gold

MIPRO 2012


This paper presents a pilot study towards estimating complex traffic flow parameters in airborne video. The study presents two prototype software systems attempting to solve intermediate tasks in recovering microscopic OD (origin-destination) matrices at complex road intersections. The first system employs background modelling in order to estimate the OD matrix of an intersection imaged by a fixed camera. The second system explores the feasibility of applying such approach to input video acquired from a hovering aircraft by pre-warping the whole video towards the coordinates of the first frame. The experimental part presents performance evaluation of the two prototype systems on real traffic videos acquired from a tall building and a non-rigid airship. The paper is concluded by discussing the achieved baseline performance and proposing suitable directions for future research.


This work was done as part of BSc project. It consisted of two distinctive but related parts: i) stabilization of airborne video and ii) traffic tracking and estimation of origin-destination matrices for given intersections. The goal was of course to be able to use UAV-as to monitor deliberate intersections and obtain turning statistics that can be further be used in traffic analysis and planning by the institutes in charge.

The first part was done by Mateja and consisted of taking a video like this (here, stable features are marked in red):

and using those features to estimate a homography and stabilize the video:

Viktor and I worked on tracking vehicles and computing statistics from an already stable video (filmed from a tall building). As always, to learn more, we didn't use and framework functions and code everything ourselves. I was working on performing gaussian blurring with MMX instructions and OpenMP, modeling the background and partially dilation and erosion. functions. Viktor fine tuned dilation and erosion, implemented vehicle tracking and made computing the origin-destination matrix on the top-right corner possible:

Of course, this work is now obsolete and nobody would use a background subtraction method to track vehicles, especially not from a less stable aerial vehicle. However, for us, it was a nice learning experience and great introduction to computer vision and its classical methods.

Full Article

Click here

View on Google Scholar

Click here

To Cite