21. June 2014

Motion Segmentation of RGB-D Videos via Trajectory Clustering

Persistence paid off. The little, remaining work on my Master’s thesis at KTH is finally completed. I can hardly remember touching a subject at university that was as demanding and as satisfying as the research on how to segment moving objects in videos taken by a RGB-D camera. Thanks to everyone who helped me with this mission.

The thesis was exploring segmentation moving objects in RGB-D videos. My previous post showed videos of model semi-trailer trucks, where the goal was to segment the moving trucks from the background. These videos were recorded in collaboration with the KTH Smart Mobility Lab. I also put together my first animated movies in the course of the thesis (using the fantastic open-source software Blender), featuring a remote-controlled Volkswagen Bus driving around in a virtual living room, see here:

The green and the blue dots in the video are landmarks in the scene that are tracked over time and then later clustered together. The groups of landmarks correspond (hopefully) to moving objects. The details can be found in the written report of my Master’s thesis. More videos can be found on my YouTube channel and in the supplementary material.

Many, many thanks to the folks at Volumental, KTH Smart Mobility Lab, my family for help and the Blender tutorial, my friends who gave feedback on the draft report, my opponent for the feedback and the tough questioning, the support from my home university, and our professor in Numerical Programming at TUM who was mindful to drill the importance of eigenvectors and eigenvalue into every student. Thanks. I learnt a lot on this journey.