A Monocular Indoor Localiser Based on an Extended Kalman Filter and Edge Images from a Convolutional Neural Network

Robot Localization in Floor Plans Using a Room Layout Edge Extraction Network

“…Instead of locally reconstructing the 3D world from camera observations and matching this reconstruction to an extruded model of the floor plan, we project the lines extracted from the floor plan into the camera frame. Our approach shares similarity with the work of Chu and Chen [15], Wang et al [16] and Unicomb et al [12]. In the first two works the authors localize a camera using a 3D model extracted from a floor plan.…”

“…Such points are then used as virtual endpoints for vanilla MCL. A similar approach by Unicomb et al [12] was proposed recently to localize a camera in a 2D map. The authors employ a CNN for floor segmentation from which they identify which lines in an edge image belong to the floor plan.…”

“…In contrast to these approaches, we use a CNN to reliably predict room layout edges in order to better cope with occlusion due to clutter and furniture. Unicomb et al [12] also employ a CNN but they only learn to extract floor edges which is a limitation in the case of clutter or occlusions. Furthermore, using a Kalman Filter approach to project the measurement into the floor plan of the map can make the system less robust to wrong initialization as the accuracy of the virtual measurement is dependent on the current camera pose estimation.…”

“…Most of the CNN-based approaches for estimating room layout edges employ a encoder-decoder topology with a standard classification network for the encoder and utilize a series of deconvolutional layers for upsampling the feature maps [4], [5], [17], [12]. Ren et al [17] proposed an architecture that employs the VGG-16 network for the encoder followed by fully-connected layers and deconvolutional layers that upsample to one quarter of the input resolution.…”