Towards scene understanding implementing the stixel world

Abstract

In this paper, we present our work towards scene understanding based on modeling the scene prior to understanding its content. We describe the environment representation model used, the Stixel World, and its benefits for compact scene representation. We show our preliminary results of its application in a diverse environment and the limitations reached in our experiments using imaging systems. We argue that this method has been developed in an ideal scenario and does not generalise well to uncommon changes in the environment. We also found that this method is sensitive to the quality of the stereo rectification and the calibration of the optics, among other parameters, which makes it time-consuming and delicate to prepare in real-time applications. We think that pixel-wise semantic segmentation techniques can address some of the shortcomings of the concept presented in a theoretical discussion.