Simple monocular door detection and tracking

Abstract

When considering an indoor navigation without using any prior knowledge of the environment, relevant landmark extraction remains an open issue for robot localization and navigation. In this paper, we consider indoor navigation along corridors. In such environments, when considering monocular cameras, doors can be seen as important landmarks. In this context, we present a new framework for door detection and tracking which exploits geometrical features of corridors. Since real-time properties are required for navigation purposes, designing solutions with a low computational complexity remains a relevant issue. The proposed algorithm relies on visual features such as lines and vanishing points that are further combined to discriminate the floor and wall planes and then to recognize doors within the image sequences. Detected doors are used to initialize a dedicated edge-based 2D door tracker. Experiments show that the framework is able to detect 82% of doors on our dataset while respecting real time constraints.