Scene Description from Depth Images for Visually Positioning

Visually positioning of impaired and blind people is still a challenge for computer vision community. The difficulty still remains in the retrieval process which try to find the best place associated to a query image. We propose in this paper a scene descriptor suitable as a filter in the matching process between query and model frames. From a video sequence of depth images, our method based geometry, represents the scene using 2D map. First, we segment the scene from the depth image into planar regions. Using these regions, the application of registration algorithm allows obtaining 3D model of the scene. Next, we determine the ground plan independently from RGB-D camera's pose and we select walls as the lateral and vertical plans to ground. Their projection on the ground gives the boundaries of the scene which are described geometrically. Finally, the descriptor of the obtained 2D map provides a determining selector of place classes. We applied our method on indoor scenes, the obtained results are presented and discussed.