Appearance-Based Indoor Navigation by IBVS Using Line Segments

Combining line segments and points for appearance-based indoor navigation by image based visual servoing

“…Most of the techniques employed in outdoor environments typically results in a significant performance drop when applied to indoor scenarios because of windows, wiry structures, reflections and repetitions, as well as limited texture in indoor scenarios, which causes standard procedures based on image descriptors to poorly perform indoors [17]. However, our previous work [9], along with SLAM based methods like [4], [5], [18], and corridor navigation methods like [19], [20], [21], showed that line segments are actually good features for indoor navigation. Still, there are some limitations mainly due to the line matching/tracking algorithms, which are still not a mature field in computer vision unlike points, especially when very few line segments can be detected in the images.…”

“…On the other hand, the appearance-based approaches do not require a 3D model of the environment, and work directly in the sensor space. To simplify the process of such navigation schemes, the environment is generally represented as topological graphs [6], [7], [8], [9]. The nodes of the graph give characteristic features or zones of the environment (locations) obtained using the sensor data, and arcs give adjacency relations between locations.…”

“…Such maps can be built in a prior offline learning phase, which is far less complex than what needed for metric maps. The comparison between the current view and the images in the map is generally based on global descriptors, like considering the whole image [10], [11], [12], color histograms [13], or image gradient [14]; or by using local descriptors, like photometric invariants [15] Stable Extremal Regions (MSER) [6], [7], [8], [16], or line segments [9].…”

“…The works in [6], [7], [8], [9], [16] have demonstrated navigation in indoor and outdoor environments using appearance based approaches from image memory. In [6], [7], [9], it is shown that accurate mapping and localization is not mandatory for visual navigation.…”

“…In [6], [7], [9], it is shown that accurate mapping and localization is not mandatory for visual navigation. The first two methods used a hybrid model for topological navigation based on visual memory in an outdoor environment where local 3D reconstruction has been used for verifying the key-point matches and automatic key-frame selection using SIFT, Multi Scale Harris, and MSER features.…”